Generative editing
English / 中文

Imagine you are the art director.

想象一下，你是艺术总监。

Your exhibition opens tomorrow.

你的展览明天开幕。

You love this glass pear, but you want to see it in ivory ceramic.

你很喜欢这只玻璃梨，但想看看它变成象牙白陶瓷的样子。

You give the editor a tiny request: change the material, keep everything else.

你给编辑器一个小小的要求：改变材质，其他一切保持不变。

Then you notice the reflection.

然后，你注意到了倒影。

Should it still look like glass?

它还应该像玻璃一样吗？

Suddenly, a three-word edit contains a whole argument about the world.

突然，一个简单的编辑指令，包含了关于世界如何运作的一整套判断。

That is our starting point.

这就是我们的起点。

We will follow this fictional artwork through image editing, exhibition design, and a moving shot.

我们将跟随这件虚构作品，经历图像编辑、展览设计，再进入一段运动镜头。

The exhibition is called AFTER RAIN.

展览名叫《雨后》，英文是 AFTER RAIN。

The pear is our recurring character.

这只梨会反复登场。

By the end, you should be able to explain a model's mechanism, defend a visual choice, and catch a failure that a beautiful preview can hide.

到课程结束，你应该能够解释模型机制、为视觉选择提供理由，并发现精美预览可能掩盖的失败。

First, look carefully at what you think must survive.

首先，仔细看看你认为必须保留的东西。

We are going to give the same object three different jobs.

我们要让同一个物体承担三种不同的任务。

First it is a technical puzzle: can its material change while its identity survives?

首先，它是一个技术难题：能否改变材质，同时保留它的身份特征？

Then it becomes an artwork: can its placement and lighting make us feel that it is fragile?

接着，它成为艺术作品：能否通过摆放和灯光，让我们感到它的脆弱？

Finally it becomes a character in time: can it disappear behind something and return as the same object?

最后，它成为时间中的角色：能否在被遮挡后重新出现，仍然是同一个物体？

Those jobs demand different judgments.

这些任务需要不同的判断。

A technically clean image may say very little.

技术上干净的图像，可能表达得很少。

An expressive poster may contain an intentional impossibility.

富有表现力的海报，可能故意包含不可能的景象。

A wonderful still may belong to a broken video.

一张精彩的静帧，可能来自一段存在问题的视频。

As its job changes, you may find yourself approving a change that you would have rejected five minutes earlier.

随着任务改变，你可能会接受一个五分钟前还会拒绝的变化。

For a few seconds, ignore the pear.

先暂时别看梨。

Look at the water beneath it.

看看它下面的水。

Now look back at the body.

现在再看回梨的主体。

If the body becomes opaque ceramic, what happens to the light that used to pass through the glass?

如果主体变成不透明的陶瓷，原先穿过玻璃的光会怎样？

The answer cannot live entirely inside the object's outline.

答案不可能完全局限在物体轮廓之内。

This comparison is a prepared classroom illustration.

这组对比是事先准备的课堂示意图。

Use it to locate the consequences of the request: transmission, highlights, and the reflection below.

用它来找出指令引发的影响：透光、高光，以及下方的倒影。

We still want the blue stem and recognizable silhouette.

我们仍然希望保留蓝色的柄和可辨认的轮廓。

But preserving every surrounding pixel could preserve the wrong optics. a successful local edit may require a carefully justified change somewhere else.

但保留周围的每一个像素，可能会保留错误的光学效果。成功的局部编辑，有时需要在别处做出有充分理由的改变。

Let us make the opening comparison more disciplined.

让我们更严谨地审视开场的对比。

The disappearance of transmitted amber light is consistent with changing transparent glass into opaque ceramic.

从透明玻璃变为不透明陶瓷后，透过的琥珀色光消失，是合理的。

A changed blue stem, by contrast, would need a separate justification because the instruction did not request that transformation.

相比之下，如果蓝色的柄也变了，就需要另外解释，因为指令并没有要求这种变化。

The reflection is more interesting.

倒影则更有意思。

A material change can require its appearance to change, while its placement should still agree with the object and the water.

材质改变可能要求倒影外观随之变化，但它的位置仍应与物体和水面一致。

We cannot classify every difference using a simple rule that change is bad.

我们不能用“变化就是坏事”这一简单规则来判断每个差异。

We need a model of the intended scene.

我们需要理解目标场景应当如何运作。

That is why the edit contract includes allowed consequences as well as invariants.

因此，编辑约定既要包含不变量，也要包含允许发生的连带变化。

Here is the route through that puzzle.

下面是我们破解这个难题的路线。

We begin inside an image editor: its representations, sampling process, and controls.

我们先进入图像编辑器内部，研究它的表示方式、采样过程和控制手段。

Then we take the art director's seat and decide what those controls should accomplish.

然后坐到艺术总监的位置，决定这些控制手段应该实现什么。

In the final technical section, the image starts moving and the preservation problem becomes harder.

在最后一个技术部分，图像开始运动，保留原有信息的问题也变得更难。

The spoken script is planned for about ninety minutes.

讲解稿按约九十分钟规划。

The marked discussions, the break, and student presentations need additional class time.

标出的讨论、课间休息和学生展示，需要额外的课堂时间。

You will see the same pear return in different situations.

你会看到同一只梨出现在不同情境中。

Each return asks us to revise our judgment, rather than learn a completely unrelated example.

每次重逢，都要求我们修正判断，而不是学习一个毫不相关的新例子。

At the end, three readings let us question who sets the goal and who decides whether it was achieved.

最后三篇阅读，将让我们追问：谁来设定目标，又由谁判断目标是否达成？

Here are three decisions I want you to be able to defend by the end.

到课程结束，我希望你能为这三类决定提供理由。

When an edit fails, which control would you change, and why?

编辑失败时，你会调整哪一种控制手段，为什么？

When two posters both look good, which belongs in this exhibition?

两张海报都很好看时，哪一张更适合这个展览？

When a video looks convincing, which moment would you inspect before accepting it?

视频看起来可信时，你会在接受它之前检查哪个时刻？

You do need a link between mechanism and consequence.

你确实需要把机制与结果联系起来。

A mask tells us where; a reference can show us what; an artistic brief tells us why the result matters.

蒙版告诉我们改哪里；参考图可以展示改成什么；艺术创作说明告诉我们结果为何重要。

Today we will practice making those links aloud.

今天，我们会练习把这些联系说清楚。

The aim is to leave with reasons you can use when the next tool has a different interface.

目标是让你带走一套理由，即使下一款工具换了界面，也能继续使用。

The pear gives us a concrete technical problem: change its material while retaining the information that makes it recognizable.

这只梨给我们一个具体的技术问题：改变材质，同时保留让人认出它的信息。

We can describe this as conditional generation under preservation constraints.

我们可以把它描述为：在保留约束下进行条件生成。

The inputs specify the change; the contract specifies what should survive.

输入指定要改变什么；编辑约定指定要保留什么。

Throughout this section, connect each mechanism to one part of that problem.

在这一部分，请把每种机制对应到这个问题的某一环节。

We begin by making our expectations explicit.

我们先把预期说清楚。

Before we open the machinery, decide what success should look like.

在打开技术黑箱之前，先确定什么才算成功。

When glass becomes ceramic, what should stay the same?

玻璃变成陶瓷时，哪些东西应该保持不变？

Which parts of the scene must change with the material?

场景的哪些部分必须随材质一起改变？

What could a model misunderstand in that request?

模型可能会如何误解这个要求？

Discuss these three questions and point to visible evidence for your choices.

请讨论这三个问题，并指出支持你们选择的可见证据。

If your partner wants to preserve a detail that you want to change, identify the artistic intention behind each answer.

如果同伴想保留的细节恰好是你想改变的，请找出两种答案背后的艺术意图。

Pause the video here.

请在这里暂停视频。

We will return to these expectations when we write the worked ceramic-edit contract.

等到我们为陶瓷编辑写出完整约定时，再回头检视这些预期。

The gallery has become a forest.

画廊变成了森林。

Let us write the edit contract.

让我们写下编辑约定。

The intended change is the environment.

我们想改变的是环境。

The pear's identity, its pedestal, and the framing should remain recognizable.

梨的身份特征、底座和构图，都应该仍然可辨认。

The ambient light and reflections are allowed to adapt to the forest.

环境光和反射则允许适应森林。

Why include that last category?

为什么要加入最后这一类？

Because preserving every visible relationship would contradict the new setting.

因为保留所有可见关系，会与新环境矛盾。

Imagine retaining a bright gallery-window reflection inside a dark forest.

想象一下，在昏暗森林里还保留着明亮画廊窗户的反光。

The model might preserve the source faithfully and still produce an implausible scene.

模型可能忠实保留了原图，却仍然生成了不合理的场景。

Before generating, separate the requested change, the invariants, and the consequences that should adapt.

生成之前，要区分所要求的变化、不变量，以及应当随之调整的连带效果。

Afterwards, inspect those categories separately.

生成之后，再分别检查这些类别。

Suppose a curator says, keep the pear exactly the same.

假设策展人说：让这只梨完全保持原样。

There are at least two meanings hiding in that sentence.

这句话至少隐藏着两种含义。

One means that a visitor should recognize the same sculpture in a new room.

一种是：观众在新房间里，仍然能认出这是同一件雕塑。

Its highlights may change with the room.

它的高光可以随着房间而变化。

The other means that selected pixel values must remain identical.

另一种是：选定像素的数值必须完全一致。

Those are different requirements.

这是两种不同的要求。

Generative conditioning can encourage recognizable identity.

生成式条件控制可以帮助保留可辨认的身份特征。

Copying source pixels or compositing can enforce exact preservation in a specified region.

复制原图像素或进行合成，可以强制指定区域完全不变。

Neither requirement automatically makes the whole image coherent: a copied region may now have the wrong lighting.

但这两种要求都不会自动保证整幅图像协调：复制过来的区域，光照可能已经不合适。

Before choosing a tool, settle the meaning of same.

选择工具之前，先确定“相同”到底是什么意思。

Read the expression as a distribution of possible edited results, given the source, instruction, and optional references.

把这个表达式理解为：给定原图、指令和可选参考信息后，可能产生的编辑结果的分布。

The symbol theta represents the model's learned parameters.

符号 theta 表示模型学习到的参数。

The vertical bar means 'given.' We are not asking for a unique answer in every case.

竖线表示“给定”。我们并不是在所有情况下都要求唯一答案。

Two different forest scenes might both satisfy the same brief.

两个不同的森林场景，都可能满足同一份创作说明。

However, a source image is a condition rather than a promise that every unmentioned pixel will be copied.

但原图只是条件，并不意味着每一个未提及的像素都会被复制。

To judge success, we need a more precise contract than 'it looks plausible.'

要判断是否成功，我们需要比“看起来合理”更精确的约定。

The probability expression says that the system produces a candidate given conditions.

概率表达式表示：系统在给定条件下生成一个候选结果。

It does not contain a certificate that the candidate passes our checks.

它并不附带证明，保证这个结果通过我们的检查。

The acceptance step is a separate decision that we impose on the result.

是否接受，是我们对结果做出的另一个独立判断。

Imagine two forest outputs.

想象两个森林版本。

Both look plausible, but one changes the pear's stem and one retains it.

两个都看起来合理，但一个改变了梨柄，另一个保留了它。

The model can assign probability to both; our contract can reject one.

模型可以给两者都分配概率，而我们的约定可以拒绝其中一个。

In practice, acceptance may involve human inspection, image comparisons, or explicit production constraints.

实际上，验收可能依赖人工检查、图像对比，或明确的制作约束。

The RGB image has 1024 by 1024 pixels and three channels: just over three million scalar values.

RGB 图像有 1024 乘 1024 个像素，每个像素有三个通道，总共略多于三百万个标量值。

With spatial reduction by sixteen and sixty-four latent channels, the representation becomes 64 by 64 by 64: 262,144 values.

空间尺寸缩小十六倍、潜在通道数为六十四时，表示变成 64 乘 64 乘 64，即 262,144 个数值。

Notice the trap.

注意这里的陷阱。

Sixteen times smaller along each spatial axis does not mean sixteen times fewer values overall, because the number of channels changes too.

每个空间轴缩小十六倍，并不意味着总数值量减少十六倍，因为通道数也变了。

Here the scalar count falls by a factor of twelve.

在这个例子中，标量数量减少了十二倍。

That is specific to this configuration.

这个比例只适用于这一具体配置。

The model works in a learned representation, and its decoder must recover visible detail.

模型在学习得到的表示中工作，解码器必须从中恢复可见细节。

Our next question is what that compression already loses before we request any edit.

接下来要问的是：在我们提出任何编辑要求之前，压缩已经丢失了什么？

Suppose the small exhibition caption is already blurry after an edit.

假设编辑后，展览的小字号说明文字变模糊了。

You might spend an hour rewriting the prompt to say, preserve every letter.

你可能花一个小时重写提示词，强调要保留每一个字母。

There is a quicker diagnostic: encode the original image and decode it again without requesting a change.

有一个更快的诊断方法：不要求任何改变，先把原图编码，再解码回来。

If the letters are already damaged in that round trip, part of the problem lies in representation and reconstruction.

如果这次往返已经损坏了文字，那么问题的一部分就在表示和重建环节。

A more emphatic instruction cannot restore a detail the editing path never retained faithfully.

再强硬的指令，也无法恢复编辑流程从未忠实保留的细节。

Look at fine text, thin edges, and small textures first.

首先检查小字、细线边缘和微小纹理。

If reconstruction is sound but the edit damages them, investigate the editing stage.

如果重建没有问题，而编辑损坏了它们，就调查编辑阶段。

It also explains why exact exhibition typography deserves its own editable layer.

这也说明，要求精确的展览文字排版，应当保留在独立的可编辑图层中。

Now consider the number of tokens a transformer processes.

现在考虑 Transformer 需要处理多少个 token。

In our toy example, one token per position on a 64 by 64 grid gives 4,096 tokens.

在这个简化例子里，64 乘 64 的网格上，每个位置对应一个 token，共有 4,096 个。

Grouping each two-by-two region gives 1,024 tokens, a reduction by four.

把每个二乘二区域组合起来，就得到 1,024 个 token，数量减少四倍。

Dense self-attention compares token pairs.

稠密自注意力会比较 token 对。

Squaring the two counts gives 16,777,216 and 1,048,576 pair scores.

将两个数量分别平方，得到 16,777,216 和 1,048,576 个配对分数。

That is a reduction by sixteen.

数量减少了十六倍。

This is why representation choices matter so much for cost.

这就是为什么表示方式会如此显著地影响成本。

It is not a claim that the whole system runs sixteen times faster.

但这不等于整个系统的运行速度会提高十六倍。

Text tokens, reference tokens, other layers, and implementation details also matter.

文本 token、参考图 token、其他网络层和实现细节，也会影响结果。

Before reading the second number, make a prediction.

看第二个数字之前，先预测一下。

We double the image width and double its height, keeping the patch scheme fixed.

我们把图像宽度和高度都加倍，同时保持分块方式不变。

How many image tokens do we get?

图像 token 会变成多少？

Four times as many.

原来的四倍。

In dense self-attention, each token can compare with every other token.

在稠密自注意力中，每个 token 都可以与其他所有 token 比较。

Four times four gives sixteen times as many pair scores.

四乘四，配对分数的数量就变成十六倍。

This calculation describes the image-token attention component, not a promise that the entire system becomes sixteen times slower.

这个计算只描述图像 token 的注意力部分，并不保证整个系统会慢十六倍。

Other operations and implementations matter.

其他操作和实现方式同样重要。

The useful habit is to ask what grew: pixels, tokens, comparisons, or measured runtime.

一个有用的习惯是追问：增加的究竟是像素、token、比较次数，还是实测运行时间？

They are related quantities, but they are not interchangeable.

这些量相互关联，但不能互换。

Let us walk through this architecture from the input side.

让我们从输入端走一遍这个架构。

The source image contributes semantic information through a vision-language component and visual information through a VAE.

原图通过视觉语言组件提供语义信息，并通过 VAE 提供视觉信息。

The target stream begins with an evolving noisy representation.

目标分支从一个不断演化的带噪表示开始。

The transformer processes that stream together with the conditions.

Transformer 将这个分支与条件信息一起处理。

During training, we possess an example of the desired edited target, so we can measure what the model should learn.

训练时，我们拥有所需编辑结果的样本，因此可以衡量模型应该学习什么。

During inference, we have the source and request but not the desired target.

推理时，我们有原图和要求，却没有所需的目标图像。

The model must generate it.

模型必须把它生成出来。

That difference is essential.

这个区别至关重要。

The architecture does not quietly receive the answer when you ask it to edit your photograph.

当你让模型编辑照片时，架构并没有偷偷收到答案。

There is a piece of information on the left that we do not possess on the right: the desired edited image.

左边有一项信息，是右边没有的：期望得到的编辑图像。

During supervised training, a source, an instruction, and a target show the model what a successful change looks like.

在监督训练中，原图、指令和目标图一起，向模型展示成功的变化是什么样子。

During use, we provide the request because that target does not yet exist.

使用时，我们之所以提出要求，正是因为目标图像还不存在。

This distinction explains a surprisingly common confusion.

这个区别解释了一个相当常见的混淆。

Showing a reference to an editor is not automatically teaching new weights.

给编辑器看一张参考图，并不自动意味着学习新的权重。

It may simply be conditioning this particular generation.

它可能只是在为这一次生成提供条件。

Adaptation, which we will reach with LoRA, changes trainable parameters.

适配则会改变可训练参数，讲到 LoRA 时我们会展开。

It is available when we construct the learning signal, and absent when we ask the trained model to make our ceramic pear.

构造学习信号时，目标图像是已知的；要求训练好的模型生成陶瓷梨时，它则是未知的。

Think of these two representations using our pear.

用这只梨来理解这两种表示。

Semantic features help with questions such as: which object is the pear, what does ceramic mean, and what does 'it' refer to in the instruction?

语义特征帮助回答：哪个物体是梨，陶瓷是什么意思，以及指令中的“它”指什么？

Visual latents provide appearance information, including shapes and textures that help keep the particular source recognizable.

视觉潜变量提供外观信息，包括形状和纹理，帮助保留原图的独特特征。

This is a conceptual distinction, not a claim that the branches have perfectly isolated responsibilities.

这是一种概念上的区分，并不是说两个分支的职责完全隔离。

Learned representations can overlap in what they encode.

学习到的表示所编码的信息可能重叠。

If an edit fails, ask whether the model misunderstood the request or lost important visual information.

编辑失败时，要问：模型是误解了要求，还是丢失了重要的视觉信息？

How do you teach a system the difference between a pear and this particular pear?

如何教会系统区分“一只梨”和“这只特定的梨”？

That question connects to OpenSubject, research I coauthored with Yexin Liu and our collaborators.

这个问题与 OpenSubject 有关，这是我与Yexin Liu及其他合作者共同完成的研究。

A video offers something valuable: repeated observations of a subject as the view changes.

视频提供了一种宝贵资源：随着视角变化，对同一主体进行反复观察。

The shared identity is a learning opportunity.

共同的身份特征提供了学习机会。

Follow the main path in the figure.

请沿着图中的主路径看。

We curate clips, verify subjects across frames, and select diverse pairs.

我们筛选视频片段，跨帧验证主体，并挑选具有多样性的图像对。

Inpainting or outpainting helps synthesize reference inputs, followed by verification.

通过图像修补或扩图来合成参考输入，再进行验证。

The corpus contains 2.5 million samples; the contribution is training data and a benchmark.

数据集包含二百五十万个样本；这项工作的贡献是训练数据和评测基准。

For our exhibition, imagine the same sculpture photographed in several rooms.

对于我们的展览，可以想象在不同房间拍摄同一件雕塑。

We want freedom to change the setting without losing its distinctive features.

我们希望自由改变环境，同时不丢失它的独特特征。

That is a classroom application of the identity problem, rather than a claim that our fictional pear was tested in the paper.

这是身份保持问题的课堂应用，并不意味着论文测试过这只虚构的梨。

The editor needs a way to learn how to move from noise toward an image.

编辑器需要学习如何从噪声走向图像。

In this simple flow-matching setup, we construct an intermediate latent by mixing noise, epsilon, with a known target latent, z one.

在这个简单的流匹配设定中，我们将噪声 epsilon 与已知目标潜变量 z₁ 混合，构造中间潜变量。

The mixing variable t tells us where we are along that training path.

混合变量 t 告诉我们处于这条训练路径的什么位置。

For this straight interpolation, the target velocity is the target latent minus the noise.

对于这种直线插值，目标速度等于目标潜变量减去噪声。

The model sees an intermediate state and learns to predict that direction, given the source and instruction.

模型看到中间状态，并在原图和指令的条件下，学习预测这个方向。

There is also an important clock distinction: t here is the generative process's time.

还要区分两种时钟：这里的 t 是生成过程中的时间。

Later, our video will have another time axis—the seconds that pass in the scene.

稍后的视频还有另一条时间轴：场景中实际流逝的秒数。

Before trusting an equation, check its easiest cases.

相信一个方程之前，先检查最简单的情况。

At t equals zero, the coefficient on the target becomes zero, so we recover the noise.

当 t 等于零时，目标的系数为零，所以我们得到噪声。

At t equals one, the coefficient on noise vanishes, so we recover the target latent.

当 t 等于一时，噪声的系数消失，所以我们得到目标潜变量。

Along this simple straight path, the difference between target and noise gives the direction we train the model to predict.

沿着这条简单直线路径，目标减去噪声，就是我们训练模型预测的方向。

You do not need to imagine a recognizable half-finished image at every intermediate point.

不必想象每个中间点都对应一张能辨认的半成品图像。

The calculation takes place in a learned representation, and this path is a teaching construction rather than the only possible design.

计算发生在学习得到的表示中；这条路径用于教学，并不是唯一可能的设计。

Here is the smallest possible sampling calculation.

这是一个最小的采样计算例子。

One coordinate currently equals 0.20.

当前某个坐标值是 0.20。

Its predicted velocity is 0.60, and the next step has size 0.10.

预测速度为 0.60，下一步的步长为 0.10。

Before looking at the result, say what operation should happen.

看结果之前，先说说应该做什么运算。

We add a small amount of motion to the current state: 0.20 plus 0.10 times 0.60, giving 0.26.

在当前状态上加一小步：0.20 加上 0.10 乘 0.60，得到 0.26。

In a real latent, many coordinates update together.

在真实的潜变量中，很多坐标会同时更新。

These numbers are invented to show an Euler step, and practical samplers can use more elaborate solvers.

这些数字是为了演示欧拉步而编造的，实际采样器可以采用更复杂的求解器。

But the sequence is now less mysterious: predict a direction, take a step, repeat, then decode.

但这个过程现在不再那么神秘：预测方向，迈出一步，重复，然后解码。

Imagine following a beautifully accurate set of directions to the wrong gallery.

想象你沿着极其精确的路线指引，却走向了错误的画廊。

Taking smaller steps will not repair the destination.

把步子迈小一点，并不能修正目的地。

The same distinction helps with generative sampling.

同样的区分也有助于理解生成采样。

A better numerical solver can follow a learned field more accurately.

更好的数值求解器，可以更准确地沿着学到的向量场前进。

That alone does not settle an ambiguous editing request.

但仅凭这一点，无法解决含糊的编辑要求。

If the pear's boundary is unstable across sampling settings, numerical or model behavior may matter.

如果不同采样设置下梨的边界不稳定，数值计算或模型行为可能有影响。

If we never specified whether the glass reflection should become ceramic, we also have a brief problem.

如果我们从未说明玻璃倒影是否应变成陶瓷倒影，那么创作说明本身也有问题。

Ask whether the system struggled to realize a clear instruction, or whether we have not agreed on what success means.

要问：系统难以实现一个清楚的指令，还是我们尚未就成功的含义达成一致？

Those situations call for different next actions.

这两种情况需要不同的下一步行动。

Where does the ability to follow an edit instruction come from?

遵循编辑指令的能力从何而来？

A training example can contain a source image, an instruction, and the corresponding edited target.

一个训练样本可以包含原图、指令，以及对应的编辑目标图像。

InstructPix2Pix is an early example that used synthetic editing data to teach this relationship.

InstructPix2Pix 是一个较早的例子，它用合成编辑数据来学习这种关系。

Suppose a training pair says 'make the object ceramic,' but its target also moves the camera and replaces the background.

假设某个训练对的指令是“把物体变成陶瓷”，但目标图还移动了相机并替换了背景。

The learning signal no longer cleanly identifies the requested transformation.

这时，学习信号就不再能清晰地指向所要求的变换。

The model may learn unwanted associations.

模型可能学到不希望出现的关联。

This gives us an important data-quality question: did the target accomplish the requested change while preserving what should remain?

因此，一个重要的数据质量问题是：目标图是否完成了要求的变化，同时保留了应该保留的内容？

Attractive targets alone are not enough to teach dependable editing.

仅有好看的目标图，还不足以教会可靠的编辑。

A training pair is a lesson, and lessons can accidentally teach the wrong thing.

一个训练对就是一堂课，而课程可能无意中教错东西。

Imagine a source glass pear and a target ceramic pear that also has a different background.

想象原图是一只玻璃梨，目标图是一只陶瓷梨，但背景也变了。

The written request says only to change the material.

文字要求却只说改变材质。

Which visible changes should the model associate with that request?

模型应当把哪些可见变化与这个要求联系起来？

If this mismatch recurs in the training data, the model can learn an unwanted association between a material change and a scene change.

如果训练数据中反复出现这种不匹配，模型可能把材质变化与场景变化错误地联系起来。

Compare the instruction with the actual difference between source and target.

请将指令与原图、目标图之间的实际差异进行比较。

Ask what supervision is rewarding, including changes nobody meant to label.

要问监督信号在奖励什么，包括那些没人打算标注的变化。

The preservation contract begins in the examples used to teach the editor.

保留约定，应当从用来教会编辑器的样本开始。

Sometimes a conditioned prediction moves in the right direction but too weakly.

有时，条件预测的方向是对的，但力度不足。

Classifier-free guidance uses the difference between a conditioned prediction and a baseline prediction to steer the result.

无分类器引导利用条件预测与基线预测之间的差异，来引导结果。

Read the formula as baseline, plus a scaled change in direction.

可以把公式读成：基线，加上经过缩放的方向变化。

When s equals one, the baseline terms cancel and we recover the conditioned prediction.

当 s 等于一时，基线项抵消，我们得到条件预测。

Above one, we extrapolate beyond it.

大于一时，我们会外推到条件预测之外。

That can strengthen the requested change, but it can strengthen errors as well.

这可以强化所要求的变化，但也可能强化错误。

In an editing system, the baseline may still retain image information; baseline does not always mean no information at all.

在编辑系统中，基线可能仍保留图像信息；基线并不总意味着完全没有信息。

Think of guidance as a specific operation on predictions, rather than a universal quality dial.

把引导理解为对预测值的一种具体运算，而不是万能的质量旋钮。

The baseline predicts 0.2 and the conditioned version predicts 0.5.

基线预测是 0.2，条件预测是 0.5。

With guidance scale two, do we get a value somewhere between them?

引导强度设为二，结果会落在两者之间吗？

No.

不会。

We take 0.2 plus twice the difference of 0.3, which gives 0.8.

我们用 0.2 加上两倍的差值 0.3，得到 0.8。

We have moved beyond the conditioned prediction.

我们已经越过了条件预测。

If the useful direction includes a slight mistake, we can amplify both.

如果有效方向中包含一点错误，那么两者都可能被放大。

For the pear, the new material might become clearer while the stem or boundary becomes less faithful.

对于这只梨，新材质可能更明显，但柄或边界可能变得不那么忠实。

Compare the requested change and preservation separately.

请分别比较所要求的变化和保留效果。

One slider can move those judgments in opposite directions; a single overall impression can hide the tradeoff.

同一个滑块可能让两项评价朝相反方向变化；单一的总体印象会掩盖这种取舍。

Different controls communicate different kinds of information.

不同控制手段传达不同类型的信息。

Text describes a requested change.

文本描述所要求的变化。

A mask specifies a region.

蒙版指定区域。

A spatial condition can describe pose, depth, or edges.

空间条件可以描述姿态、深度或边缘。

An image reference can supply appearance information that would be difficult to express precisely in words.

参考图可以提供难以用语言精确表达的外观信息。

These are not interchangeable knobs, and every product does not expose all of them.

这些控制并不能互相替代，也不是每款产品都提供全部选项。

ControlNet and IP-Adapter are examples of distinct architectural approaches.

ControlNet 和 IP-Adapter 就是两种不同架构思路的例子。

If exact untouched pixels are essential, a generation mask alone may be insufficient; explicit copying or compositing can enforce that requirement.

如果必须精确保留未编辑像素，仅有生成蒙版可能不够；明确复制像素或合成，可以强制实现这一要求。

Choose a control by asking what information is missing, then check whether the resulting image actually respected it.

选择控制手段时，先问缺少什么信息，再检查结果是否真正遵循了它。

A mask indicates a permitted region; it does not by itself solve seams or the reflection of a changed material.

蒙版指出允许编辑的区域，但它本身不能解决接缝，也不能处理材质改变后的倒影。

The next slide names an architecture that learns to use a spatial condition.

下一页介绍一种学习使用空间条件的架构。

Suppose your sentence is understood, but the pear keeps changing shape.

假设模型理解了你的话，但梨的形状总在变化。

You can describe its outline with more adjectives, or supply spatial evidence.

你可以用更多形容词描述轮廓，也可以直接提供空间证据。

ControlNet gives an edge map, depth map, or pose a learned route into a compatible diffusion model.

ControlNet 为边缘图、深度图或姿态信息，提供了一条进入兼容扩散模型的可学习路径。

Trace the two paths in this original architecture.

请追踪原始架构中的两条路径。

The pretrained path stays frozen.

预训练主路径保持冻结。

A trainable copy processes the added condition, and zero-initialized one-by-one convolutions connect its features to the main path.

一个可训练副本处理新增条件，零初始化的一乘一卷积把它的特征连接到主路径。

Initially those connections contribute zero; training learns their contribution.

初始时，这些连接的贡献为零；训练会学出它们应有的贡献。

The copied branch itself starts from pretrained weights, not all zeros.

副本分支本身从预训练权重开始，而不是全部从零开始。

Now return to our request.

现在回到我们的要求。

Edges can help specify the outline.

边缘可以帮助指定轮廓。

They cannot, by themselves, tell us whether the reflected ceramic looks convincing or the blue stem retains its identity.

但仅凭边缘，无法判断陶瓷倒影是否可信，也无法保证蓝色梨柄保留身份特征。

With several references, the system must know which reference contributes which information.

当有多张参考图时，系统必须知道每张图分别提供什么信息。

Imagine asking for the sculpture from image one and the atmosphere from image two.

想象你要求使用图一的雕塑，以及图二的氛围。

If the references become confused, you may get the wrong object with the right lighting.

如果参考图的角色混淆，可能会得到光照正确、物体却错误的结果。

The illustrated research approach uses separators and image-index information to distinguish inputs.

图示研究方法使用分隔符和图像索引信息来区分输入。

In an art brief, name the role of each image rather than presenting a pile of vaguely related inspiration.

在艺术创作说明中，要说明每张图的角色，而不是堆放一批关系模糊的灵感图。

Then inspect for leakage: did a composition reference accidentally replace the subject?

然后检查信息是否串用：构图参考是否意外替换了主体？

This figure describes one proposed mechanism, not a universal design used by every editor.

这张图描述的是一种提出的机制，并不是所有编辑器通用的设计。

Look at these published multi-image examples with reference roles in mind.

请带着“参考图角色”的概念，观察这些已发表的多图示例。

Before judging the output, identify what each input was supposed to contribute.

评价输出之前，先确定每个输入原本应该贡献什么。

Then trace the subject and the requested transformation into the result.

然后在结果中追踪主体和所要求的变换。

A reference-role failure can look superficially attractive.

参考图角色混淆的结果，表面上也可能很好看。

The system may borrow the wrong object's appearance or import a background that was never intended.

系统可能借用了错误物体的外观，或引入了原本不需要的背景。

We are examining qualitative examples from a particular paper, not conducting a broad comparison between products.

我们正在查看某篇论文的定性示例，并不是对产品做全面比较。

The figure helps us practice an inspection method: follow the intended contribution of each input and look for unintended transfers between them.

这张图帮助我们练习一种检查方法：追踪每个输入的预期贡献，寻找意外的信息转移。

Attention lets a token gather information from other tokens with different weights.

注意力让一个 token 以不同权重，从其他 token 收集信息。

Our simplified scalar example gives weight 0.8 to a value of 0.9, and weight 0.2 to a value of 0.1.

在这个简化的标量例子中，数值 0.9 的权重为 0.8，数值 0.1 的权重为 0.2。

The weighted result is 0.74.

加权结果是 0.74。

The real mechanism operates on learned vectors, so this is arithmetic intuition rather than a literal description of artistic decision-making.

真实机制处理的是学习到的向量，因此这里只是在建立算术直觉，并非字面描述艺术决策。

The important structure is selective combination: every available source need not contribute equally.

关键结构是选择性组合：不是每个可用来源都必须贡献同样多的信息。

In a multi-reference edit, the system also needs to distinguish where information came from.

在多参考图编辑中，系统还需要区分信息来自哪里。

Otherwise, gathering information successfully can still produce the wrong mixture of subject identity and visual style.

否则，即使成功收集了信息，也可能错误地混合主体身份与视觉风格。

The weights in this simplified attention example sum to one.

在这个简化注意力例子中，权重之和为一。

That makes the result a weighted combination of the values.

因此，结果是各个数值的加权组合。

A larger weight makes the associated value contribute more to this particular calculation.

权重越大，对应数值在这次计算中的贡献就越大。

Now be cautious about jumping from the calculation to an explanation of the final image.

但要谨慎，不能直接从这个计算跳到对最终图像的解释。

Real networks contain many layers, heads, and transformations.

真实网络包含许多层、注意力头和变换。

A high weight at one point may not tell us why a visible feature ultimately appeared.

某处权重很高，未必能解释某个可见特征最终为何出现。

For practical reference editing, the useful question remains whether the intended information survives in the output, regardless of how compelling an attention visualization looks.

在实际参考图编辑中，关键仍是预期信息是否出现在输出中，无论注意力可视化看起来多么有说服力。

What if a useful concept must recur across many requests?

如果一个有用的概念需要在许多请求中反复出现，怎么办？

A reference can condition a generation, while LoRA adapts selected learned weights using a low-rank update.

参考图可以为一次生成提供条件；LoRA 则通过低秩更新，调整选定的模型权重。

The update is a product of two smaller matrices.

更新量是两个较小矩阵的乘积。

Instead of learning every entry of a large update matrix independently, we express that update as a product of two smaller matrices.

我们不独立学习大型更新矩阵的每个元素，而是用两个较小矩阵的乘积来表示更新。

For a 4096 by 4096 weight matrix and rank sixteen, the full matrix has 16,777,216 entries.

对于一个 4096 乘 4096、秩设为十六的例子，完整矩阵有 16,777,216 个元素。

The two factors together have 131,072, a factor of 128 fewer for this update.

两个因子合计只有 131,072 个元素，这次更新的参数量减少了一百二十八倍。

That is not a claim that the entire model shrinks by 128.

这并不意味着整个模型缩小了一百二十八倍。

The base weights still exist.

基础权重仍然存在。

Low rank makes adaptation economical, but does not certify that the learned subject survives a new pose.

低秩让适配更经济，但不能证明学到的主体在新姿态下仍能保持一致。

That question requires examples the adaptation did not already see.

这个问题需要用适配时未见过的样例来检验。

An adapted model reproduces your favorite training portrait perfectly.

适配后的模型完美复现了你最喜欢的训练肖像。

Is that enough to use the character in a new film?

这就足以让这个角色出演新电影了吗？

Consider what else it may have learned: the familiar camera angle, the background, even the lighting that always accompanied the subject.

想一想它还可能学到了什么：熟悉的机位、背景，甚至总与主体一起出现的光照。

A held-out check deliberately changes those circumstances.

留出测试会有意改变这些条件。

Ask for a new view or a different setting and inspect the distinctive features.

要求一个新视角或不同环境，并检查独特特征。

We want a reusable concept, rather than a narrow ability to reproduce familiar combinations.

我们需要的是可复用的概念，而不只是复现熟悉组合的狭窄能力。

For the pear, keep the blue stem recognizable while moving the exhibition outdoors.

对于这只梨，要在把展览移到户外时，仍保留可辨认的蓝色柄。

If identity collapses there, more praise for the training examples will not solve the problem.

如果身份特征在那里崩溃，再称赞训练样例也解决不了问题。

A reward tells a learning process what kinds of outputs to favor.

奖励告诉学习过程应当偏好哪些输出。

The figure shows a recent approach that separates task-specific reward training and then distills what is learned.

图中是一种近期方法：分别训练任务特定的奖励，再蒸馏学到的能力。

Look at the categories: editing quality involves more than a single judgment of attractiveness.

看这些类别：编辑质量不只是单一的美观判断。

Now imagine an artwork whose purpose is to feel awkward or disturbing.

现在想象一件艺术作品，其目的就是让人感到别扭或不安。

A generic preference for polished images could work against that purpose.

对精致图像的普遍偏好，可能反而违背这个目的。

Human preference signals are useful, but they do not define artistic merit for every project.

人类偏好信号很有用，但不能为所有项目定义艺术价值。

This is the bridge to the next section: technical optimization can help produce candidates, while an artist still needs to decide which properties serve the work.

这正是通向下一部分的桥梁：技术优化可以帮助生成候选作品，而艺术家仍要决定哪些属性服务于作品。

These edit examples invite a question about evaluation.

这些编辑示例引出了一个评价问题。

Which properties would you reward separately?

你会分别奖励哪些属性？

You might ask whether the instruction was carried out, whether identity survived, and whether the image is visually convincing.

可以问：指令是否完成、身份是否保留，以及图像在视觉上是否可信。

Those judgments can disagree.

这些判断可能彼此冲突。

A highly polished output might erase an awkward feature that is central to the artwork.

高度精致的输出，可能抹掉作品中至关重要的别扭特征。

An unusual composition might serve the brief while attracting a lower generic preference score.

不寻常的构图可能符合创作要求，却得到较低的通用偏好分数。

We should understand what a reward encourages before treating a high score as artistic approval.

在把高分视为艺术认可之前，我们应当理解奖励究竟鼓励什么。

The published examples illustrate the research setting; our classroom task is to articulate the intention against which we would judge a particular result.

已发表示例展示的是研究情境；课堂任务则是明确创作意图，并据此判断具体结果。

Which failure would you notice first?

你会先注意到哪种失败？

On one side, the output is beautiful, but it has quietly replaced our particular sculpture with a generic decorative pear.

一边的结果很美，却悄悄把我们的特定雕塑换成了普通的装饰梨。

On the other, the right sculpture is present, but its reflection still behaves like the old material.

另一边保留了正确雕塑，但倒影仍像原来的材质。

The first may win a quick aesthetic vote.

第一种可能赢得快速的审美投票。

The second may pass a checklist that only asks whether the requested object changed.

第二种可能通过只检查“目标物体是否改变”的清单。

Neither satisfies the full brief.

两者都没有满足完整要求。

Has the editor lost identity, broken the scene's optical logic, or missed the work's intention?

编辑器是丢失了身份、破坏了场景光学逻辑，还是偏离了作品意图？

Naming the mismatch gives us a next action.

说清不匹配之处，才能决定下一步。

Saying only that an output looks a little wrong leaves the diagnosis unfinished.

只说输出“看起来有点不对”，诊断还没有完成。

A diagnosis becomes useful when it changes the next action.

只有能改变下一步行动的诊断，才有用。

If the wrong object changes, first suspect an ambiguous reference or region specification.

如果改错了物体，首先怀疑参考对象或区域描述有歧义。

If the correct object becomes another instance, appearance grounding may be the issue.

如果改对了物体，却变成另一个个体，问题可能在外观信息的约束。

If the material looks disconnected from its reflection, the dependent change may be missing.

如果材质与倒影脱节，可能缺少必要的连带变化。

These are candidate explanations, not automatic conclusions.

这些只是候选解释，不是自动成立的结论。

Choose a small revision that tests one of them.

选择一个小调整，检验其中一种解释。

If it does not help, reconsider the hypothesis.

如果没有帮助，就重新考虑假设。

This is more informative than changing the prompt, seed, reference, and model all at once, because then a successful result would leave us unsure which intervention mattered.

这比同时改提示词、随机种子、参考图和模型更有信息量；否则，即使成功，也不知道哪个改动起了作用。

We now have several ways to communicate an edit.

现在，我们有了几种表达编辑要求的方法。

Would you preserve the reflection or let it change, and why?

你会保留倒影，还是允许它变化？为什么？

When would a mask help more than a longer prompt?

什么时候蒙版比更长的提示词更有帮助？

What would edges or depth control still leave uncertain?

边缘或深度控制仍会留下哪些不确定性？

Discuss these three questions using the pear already on screen.

请用屏幕上的这只梨，讨论这三个问题。

For each control you favor, name the failure it addresses and something it cannot settle.

对于你支持的每种控制手段，指出它解决什么失败，以及它不能决定什么。

Pause the video here.

请在这里暂停视频。

The next slide offers one possible contract; compare it with your reasoning rather than treating it as the only artistic answer.

下一页给出一种可能的约定；请与你的推理比较，不要把它当成唯一的艺术答案。

Here is one defensible answer to the ceramic puzzle.

对于陶瓷难题，这是一种有理有据的答案。

Preserve the blue stem and recognizable silhouette.

保留蓝色柄和可辨认的轮廓。

Allow transmission, highlights, and the reflection to change with the new material.

允许透光、高光和倒影随着新材质改变。

Then inspect the boundary and water, because that is where the request reaches beyond the object.

然后检查边界和水面，因为要求的影响会在那里超出物体本身。

Notice the wording: allow a justified consequence, rather than permit arbitrary drift.

注意措辞：允许有理由的连带变化，而不是允许任意漂移。

We have not given the model permission to redesign the gallery.

我们没有授权模型重新设计画廊。

We have named the changes necessary to make this material transformation coherent.

我们明确指出了，让这次材质变换协调一致所必需的变化。

A mask could help localize work, while compositing might protect a region that truly must remain exact.

蒙版可以帮助限定工作区域；合成则可以保护真正必须完全不变的区域。

The contract tells us how to use those tools.

编辑约定告诉我们如何使用这些工具。

We can now explain our opening puzzle at three levels.

现在，我们可以在三个层面解释开场难题。

The brief decides what should change and what should survive.

创作说明决定什么应改变、什么应保留。

The model uses representations, conditioning, and sampling to propose a result.

模型通过表示、条件控制和采样，提出候选结果。

Our review checks whether that proposal actually satisfies the brief.

我们的审查检验这个候选结果是否真正满足要求。

Keep those levels separate when you diagnose a failure.

诊断失败时，要区分这三个层面。

Stronger guidance will not choose the exhibition's purpose.

更强的引导不会替你选择展览的目的。

A more eloquent artistic statement will not enforce identical pixels.

更动人的艺术陈述不会强制像素完全一致。

A beautiful preview will not prove preservation.

精美预览不会证明保留成功。

The useful skill is connecting the right intervention to the observed problem.

有用的能力，是把正确的干预对应到观察到的问题。

If your partner can tell when to use it and what it leaves uncertain, you have understood more than its name.

如果同伴能说清何时使用它、它还留下什么不确定性，你就不只是记住了它的名字。

We can now explain our controls rather than just name them.

现在，我们能够解释控制手段，而不只是说出名称。

How does ControlNet differ from a mask or a text prompt?

ControlNet 与蒙版或文本提示词有什么区别？

Why can stronger guidance make an edit worse?

为什么更强的引导可能让编辑变差？

What evidence would show that an edit preserved identity?

什么证据能够表明编辑保留了身份特征？

Discuss these three questions.

请讨论这三个问题。

Choose a concrete requirement—a silhouette, an untouched region, or a distinctive stem—and connect your explanation to it.

选一个具体要求——轮廓、未编辑区域，或独特的梨柄——并把解释与它联系起来。

Then consider whether the control guarantees that requirement or only helps express it.

然后考虑：这种控制是保证实现要求，还是仅仅帮助表达要求？

Pause the video here before the break.

休息之前，请在这里暂停视频。

We will take ten minutes and resume when the class is ready.

我们休息十分钟，等大家准备好再继续。

When you come back, think about a series of images you would recognize as belonging to one artwork.

回来时，请想一组你能认出属于同一件艺术作品的图像。

What makes them belong together: the subject, the palette, the treatment of space, or something else?

是什么让它们属于一体：主体、色彩、空间处理，还是别的因素？

Leave that question open for now.

暂时先保留这个问题。

We will use it to move from individual editing operations to a coherent visual language.

我们将用它，从单次编辑操作转向连贯的视觉语言。

Look again at the artwork we have been using as a technical test.

再看看这件一直被我们用来做技术测试的作品。

After the break, it has a different job.

休息之后，它有了不同的任务。

We are no longer asking only whether the edit obeys a request.

我们不再只问编辑是否遵循了要求。

We are asking what a visitor might feel, and which visual decisions create that feeling.

我们要问：观众可能感受到什么，哪些视觉决定创造了这种感受？

Two images can have different surfaces and still belong to one exhibition.

两张图像的表面可以不同，却仍属于同一个展览。

Two images can share a palette and feel unrelated.

两张图像可以使用相同色彩，却感觉毫不相关。

The difference is worth arguing about.

这种差别值得讨论。

In the next section, I will show prepared alternatives for AFTER RAIN.

下一部分，我会展示为《雨后》准备的不同方案。

Choose a direction in your mind, and be ready to explain a visible reason.

请在心里选择一个方向，并准备说出可见的理由。

Your neighbor may choose the other one.

你的邻座可能会选另一个。

Now take the art director’s seat.

现在，请坐到艺术总监的位置。

The model can produce many plausible outputs; we decide which differences matter for AFTER RAIN.

模型能生成许多合理的结果；我们来决定哪些差异对《雨后》重要。

As we compare the prepared images, name the visible relationship that carries the idea.

比较这些准备好的图像时，请指出承载创意的可见关系。

That could be scale, light, or the treatment of space.

它可能是尺度、光线，或空间处理。

Your explanation will be more useful than simply calling an image impressive.

这样的解释，比简单地说图像很惊艳更有用。

Before seeing the poster alternatives, consider what you want a visitor to experience.

看海报方案之前，先想想你希望观众经历什么。

What makes an image feel fragile rather than merely attractive?

什么让图像显得脆弱，而不仅仅是好看？

Can very different images belong to the same artwork, and why?

差别很大的图像，能否属于同一件作品？为什么？

Which artistic decisions would you keep for yourself?

哪些艺术决定，你会留给自己？

Discuss these three questions.

请讨论这三个问题。

You might disagree about the effect of empty space or the meaning of a material.

你们可能对留白的效果或材质的含义有不同看法。

Locate the visual evidence behind that disagreement.

请找出分歧背后的视觉证据。

Pause the video here.

请在这里暂停视频。

Keep your starting position in mind when we compare the prepared directions.

比较预先准备的方向时，请记住你的初始立场。

Here is our commission: AFTER RAIN, a fictional exhibition about fragile objects in changing environments.

这是我们的委托：《雨后》，一个关于变化环境中脆弱物体的虚构展览。

Imagine a visitor seeing its poster across a corridor before encountering a five-second moving image inside.

想象观众先在走廊另一端看到海报，再进入展厅看到一段五秒的动态图像。

What should that visitor expect to feel?

你希望这位观众期待怎样的感受？

The prepared examples keep the amber pear and blue stem recognizable.

准备好的示例都保留了可辨认的琥珀色梨和蓝色柄。

We will examine photographic and collage directions, then consider how a moving reveal changes the experience.

我们将研究摄影和拼贴两个方向，再考虑运动中的揭示如何改变体验。

Nobody needs to generate an image during this lecture.

这节课不需要任何人现场生成图像。

Fragile is a useful beginning, but it is not yet an art direction.

“脆弱”是一个有用的起点，但还不算艺术方向。

We need to translate that word into something visible enough to compare and precise enough to revise.

我们需要把这个词转化为足够可见、能够比较，又足够精确、能够修改的东西。

Try replacing fragile with expensive in the brief.

试着把创作说明中的“脆弱”换成“昂贵”。

You might still choose glass, but would you use the same composition?

你可能仍然选择玻璃，但还会用同样的构图吗？

A large centered object and assertive lighting might suggest a luxury product.

居中放大的物体配合强势光照，可能让人联想到奢侈品。

A small object surrounded by quiet space might instead seem exposed.

安静空间中一个小小的物体，则可能显得无所庇护。

Scale can make the pear feel vulnerable.

尺度可以让梨显得易受伤害。

Restrained light can make us look closely.

克制的光线可以引导我们细看。

A reflection can suggest a world less stable than the object itself.

倒影可以暗示一个比物体本身更不稳定的世界。

These are artistic hypotheses to test against an audience's reading, not a formula that makes every image fragile.

这些是需要通过观众解读来检验的艺术假设，不是让所有图像显得脆弱的公式。

Point to the feature doing the work.

请指出真正起作用的特征。

If nobody can locate it in the image, the intention may still be living only in the prompt.

如果没人能在图像中找到它，意图可能仍然只存在于提示词里。

In Hokusai's Great Wave, look first for Mount Fuji.

看葛饰北斋的《神奈川冲浪里》，先找富士山。

It is small and distant.

它很小，也很远。

Now follow the curves of the wave and the boats beneath it.

现在沿着巨浪的曲线，以及下方的小船看。

Scale and rhythm create a tension we can discuss without turning the artwork into a style label.

尺度和节奏创造了张力；我们可以讨论这种张力，而不是把作品变成一个风格标签。

For AFTER RAIN, we might borrow a relationship: a small, vulnerable form facing a much larger environment.

对《雨后》来说，我们可以借鉴一种关系：一个小而脆弱的形体，面对远大于自身的环境。

We do not need to reproduce the wave or ask for a generic imitation of the artist.

不需要复制巨浪，也不需要要求泛泛地模仿这位艺术家。

The reference becomes useful when we can say which visual decision we are studying.

当我们能够说明正在研究哪个视觉决定时，参考作品才真正有用。

Then we must test its translation.

然后，还必须检验这种转化。

Our quiet flooded gallery has a different subject and emotional register.

我们安静而积水的画廊，有着不同的主体和情感基调。

Instead of using a reference as a label, describe a relationship you can observe.

不要把参考作品当作标签，而要描述你能观察到的关系。

You might want a small stable form beneath a large dynamic curve, or a rhythm that moves the eye through the composition.

你可能想要一个巨大动态曲线下的小型稳定形体，或引导视线穿过构图的节奏。

Then translate that relationship into the new subject and brief.

然后，把这种关系转化到新的主体和创作要求中。

The result should be judged as its own artwork, not as a contest to resemble the reference.

应把结果作为独立作品评价，而不是比赛谁更像参考图。

This approach makes references more useful to collaborators because they can understand what you are borrowing.

这种方法让参考资料对合作者更有用，因为他们能理解你借鉴的是什么。

It also makes iteration more focused: if the intended tension is missing, you can revise scale or rhythm rather than vaguely asking for more influence from the source.

它也让迭代更聚焦：如果缺少预期张力，可以调整尺度或节奏，而不是含糊地要求“更受原作影响”。

Before admiring the surface detail, decide where your eye arrives.

欣赏表面细节之前，先判断你的视线最先落在哪里。

Is it the pear, the reflection, or the space waiting above them?

是梨、倒影，还是它们上方留出的空间？

A poster has to organize that first encounter.

海报必须组织这第一次相遇。

If every region is equally busy, the audience has to invent a hierarchy we have not provided.

如果每个区域都同样繁忙，观众就得自行建立我们没有提供的层次。

Here, the open area can make the object feel small and give the title somewhere to live.

这里的开放区域能让物体显得小，也给标题留下位置。

We can ruin the sense of quiet by filling every available gap with decorative detail, even if each addition is attractive on its own.

如果用装饰细节填满每一处空隙，就可能破坏安静感，即使每个新增细节单独看都很漂亮。

When revising this image, I would first ask whether the composition expresses the intended scale and silence.

修改这张图时，我会先问：构图是否表达了预期的尺度感与寂静？

More texture comes later, if the work needs it.

如果作品需要，之后再增加纹理。

Negative space does two jobs in this poster.

留白在这张海报中承担两项任务。

It changes how large or isolated the object feels, and it creates a place for typography.

它改变物体显得多大、多孤立，也为文字排版创造位置。

These are connected design decisions rather than separate finishing steps.

这是相互关联的设计决定，而不是彼此独立的收尾步骤。

If you fill the pale area with dramatic detail, the image may become more visually active but leave no calm route for reading the title.

如果在浅色区域填满戏剧性细节，画面可能更活跃，却没有平静的路径让人阅读标题。

If you reserve too much space, the subject might lose the presence you wanted.

如果留白过多，主体又可能失去你想要的存在感。

Evaluate the space in relation to the finished use.

请结合最终用途来评价空间。

A generative background is part of a larger composition when it must carry text, branding, or other exact elements.

当背景需要承载文字、品牌或其他精确元素时，生成背景只是更大构图的一部分。

The collage direction changes the rules of the world.

拼贴方向改变了这个世界的规则。

Torn edges replace smooth contours.

撕裂边缘取代了平滑轮廓。

Flat layers replace optical depth.

平面层次取代了光学纵深。

Amber and indigo keep a connection to our recurring subject.

琥珀色和靛蓝色，让它仍与反复出现的主体保持联系。

Look at how those decisions affect the pear's apparent weight and vulnerability.

看看这些决定如何影响梨看起来的重量与脆弱感。

If the body looks like paper but the water behaves like a photograph, do we accept that collision?

如果主体像纸，水却像照片，我们是否接受这种碰撞？

We might, if it is deliberate and supports the work.

如果它是有意的，并服务于作品，我们可能接受。

We might reject it as an unresolved mixture.

我们也可能认为它是尚未解决的混杂，因而拒绝。

The answer depends on the intended visual language.

答案取决于预期的视觉语言。

The next comparison asks you to locate precisely where that language holds together or breaks.

下一组比较，请你精确指出：这种语言在哪里成立，又在哪里断裂。

A photographic reflection inside a paper collage can be a mistake, or the most interesting decision in the image.

纸质拼贴中的摄影式倒影，可能是错误，也可能是整张图最有意思的决定。

We need to know whether the mismatch is an intentional disruption and whether it produces the intended effect.

我们需要知道：这种不匹配是否是有意的打破，以及是否产生了预期效果。

Suppose the exhibition explores unstable memories.

假设展览探索的是不稳定的记忆。

An impossibly photographic reflection could make sense.

一个不可能的摄影式倒影，就可能说得通。

Suppose the brief calls for a coherent world assembled from torn paper.

假设要求是构建一个由撕纸组成的统一世界。

The same reflection might weaken it.

同样的倒影可能削弱它。

This does not mean every accident deserves an explanation after the fact.

这并不意味着每个意外都值得事后找理由。

Make the intention specific, examine what the viewer can actually see, and compare alternatives.

请明确意图，检查观众实际能看见什么，并比较不同方案。

A strong critique can distinguish a productive contradiction from an excuse for an unresolved result.

有力的评论能区分富有成效的矛盾，与为未完成结果寻找的借口。

A reference becomes easier to use when we assign it a role.

为参考图分配角色后，它就更容易使用。

One image might define the subject, another the composition, and another the material treatment.

一张图可以定义主体，另一张定义构图，再一张定义材质处理。

Name the property you want from each.

请说清你想从每张图中得到什么属性。

The model may not isolate those roles perfectly.

模型未必能完美隔离这些角色。

That is why you should look for unwanted transfers, such as importing a reference's background when you only wanted its palette.

因此，要检查意外的信息转移，例如只想借用配色，却把参考图的背景也引入了。

Try removing one reference and observing what disappears.

试着移除一张参考图，观察什么随之消失。

Start with the smallest useful set.

从最小但有用的参考集合开始。

If five references contradict one another, adding a sixth may make the problem harder to diagnose rather than solve it.

如果五张参考图互相矛盾，再加第六张可能让问题更难诊断，而不是解决问题。

We have added a reference because we think it helps.

我们添加参考图，是因为认为它有帮助。

How would we discover whether it actually does?

怎样才能知道它是否真的有帮助？

Remove it and compare.

移除它，再比较。

First state its intended role: perhaps the pear's silhouette, perhaps the composition's sense of scale.

先说明它的预期作用：可能是梨的轮廓，也可能是构图的尺度感。

Then look for that quality in outputs with and without the reference.

然后，在有无这张参考图的输出中寻找这种品质。

Also inspect what came along uninvited.

也要检查哪些信息不请自来。

A useful subject reference can carry a background or lighting scheme we did not want.

有用的主体参考，也可能携带我们不想要的背景或光照。

This is an ablation: change one component to understand its contribution.

这就是消融实验：改变一个组件，以理解它的贡献。

Random generation makes one lucky comparison weak evidence, so several samples can help.

随机生成会让一次幸运的对比缺乏说服力，因此多个样本会有帮助。

If we cannot explain what a reference contributes, we may be making the request more complicated without making the direction clearer.

如果说不清一张参考图的贡献，我们可能只是把要求变复杂，却没有让方向更清楚。

Listen to how this prompt distributes responsibility.

听听这个提示词如何分配职责。

The pear reference supplies silhouette and the blue stem.

梨的参考图提供轮廓和蓝色柄。

Placement low on the right organizes the composition.

将主体放在右下方，用来组织构图。

A pale upper-left area reserves space for the title.

左上方的浅色区域，为标题留出空间。

Soft daylight and restrained reflections support the mood.

柔和日光与克制的倒影，支撑整体情绪。

Every phrase points toward something we could inspect in an output.

每个短语，都指向输出中可以检查的东西。

Compare that with asking for a stunning masterpiece with beautiful lighting.

与“用漂亮光照创作一幅惊艳杰作”这样的要求比较一下。

This prompt is still a proposal, not a binding contract enforced by the model.

这个提示词仍然只是提议，不是模型强制执行的约束协议。

If an output fails, we can now name the failed relationship: the title area is crowded, the light is too assertive, or the reference's background has leaked into the scene.

输出失败时，我们现在可以指出哪种关系失败了：标题区太拥挤、光线太强势，或参考图背景渗入了场景。

A single poster can look coherent by itself.

单张海报本身可能看起来很统一。

A series exposes a harder problem: does the same material decision survive across viewpoints?

系列作品暴露出更难的问题：同样的材质决定，能否在不同视角下保持？

Our Group Editing collaboration studies related images that should be edited consistently.

我们合作开展的 Group Editing 研究，关注应当保持一致编辑的关联图像。

Treating each image independently can produce slightly different costumes or materials.

独立处理每张图，可能生成略有不同的服装或材质。

The method arranges related images as pseudo-video frames to use a video model's consistency prior.

该方法把相关图像排列成伪视频帧，利用视频模型的一致性先验。

VGGT provides geometric correspondences.

VGGT 提供几何对应关系。

Geometry-enhanced rotary positional embeddings connect geometry features with image latents, while Identity-RoPE supports identity preservation.

几何增强的旋转位置编码，将几何特征与图像潜变量联系起来；Identity-RoPE 则支持身份保持。

Follow the penguin examples through the figure before reading every label.

阅读每个标签之前，先沿着图中的企鹅示例看。

Reliable correspondence is part of the technical problem: when views cannot be matched well, repeating the same instruction alone does not establish a coherent series.

可靠对应关系是技术问题的一部分：视角无法准确匹配时，仅重复相同指令，并不能保证系列统一。

Iteration becomes informative when we know what changed between attempts.

当我们知道两次尝试之间改了什么，迭代才有信息量。

Suppose we compare two lighting treatments.

假设我们比较两种灯光处理。

Keep the subject and composition instructions stable, and record the references and model version.

保持主体和构图指令不变，并记录参考图和模型版本。

If a seed control exists, holding it fixed can help the first comparison.

如果可以控制随机种子，固定种子有助于初步比较。

A shared seed still does not guarantee identical composition after a prompt change.

但即使种子相同，改变提示词后，也不能保证构图完全一致。

If results vary widely, examine several outputs for each condition.

如果结果差异很大，就检查每种条件下的多个输出。

Otherwise, one lucky sample may decide the whole direction.

否则，一个幸运样本就可能决定整个方向。

The table is a proposed experiment, not measured results.

这张表是拟议实验，不是实测结果。

Its purpose is to make each generation answer a question about the artwork.

它的目的，是让每次生成都回答一个关于作品的问题。

Imagine you compare two prompts once, and the second gives a wonderful poster.

想象你只比较了两条提示词各一次，第二条生成了精彩海报。

Did its wording cause the improvement, or did it receive a favorable random sample?

改进是措辞造成的，还是它碰巧抽到了有利的随机样本？

From one output, it can be difficult to tell.

只看一个输出，往往很难判断。

Several outputs per condition reveal whether the direction is stable or merely fortunate.

每种条件生成多个结果，才能看出这个方向是稳定有效，还是仅仅幸运。

For art direction, we can ask two different questions.

在艺术指导中，我们可以问两个不同的问题。

Would I exhibit this particular image?

我愿意展出这张具体图像吗？

Could I reliably develop a series using this direction?

我能用这个方向，可靠地发展出一个系列吗？

One exceptional output may answer the first while leaving the second unresolved.

一张出色的输出可能回答了第一个问题，却没有解决第二个。

Repeating generations is helpful when it reveals a pattern relevant to the brief; it becomes a distraction when we keep browsing alternatives to avoid deciding what we value.

重复生成若能揭示与创作要求相关的规律，就有帮助；若只是为了逃避价值判断而不断浏览方案，就成了干扰。

Now the image becomes a poster.

现在，图像变成了海报。

The title is an editable typographic layer, which lets us choose its wording, line breaks, and placement exactly.

标题是可编辑的文字图层，让我们精确选择措辞、换行和位置。

That matters when a work has to carry an exhibition name rather than merely resemble a poster in a generated preview.

作品需要准确呈现展览名称时，这很重要，而不是只在生成预览中“看起来像海报”。

Look at how the title occupies the area we deliberately left open.

看看标题如何占据我们有意留出的区域。

The image and type were designed to cooperate.

图像与文字是协同设计的。

We can adjust hierarchy without asking the model to regenerate the sculpture and risk changing it.

我们可以调整层次，而不必让模型重新生成雕塑、冒着改变它的风险。

This is a useful division of labor: generation develops the visual material, and direct layout controls exact communication.

这是一种有用的分工：生成负责发展视觉素材，直接排版负责精确传达信息。

The audience sees one composition.

观众看到的是一个整体构图。

It does not need to know which parts came from which tool.

他们不需要知道哪个部分来自哪种工具。

Typography establishes a sequence of attention.

文字排版建立了注意力的顺序。

Ask what the viewer should read first, what comes next, and where their eye returns to the artwork.

要问：观众应先读什么、接着读什么，以及视线在哪里回到作品？

In our poster, reserved space allows the title to be clear without covering the sculpture.

在我们的海报中，预留空间让标题清楚可读，又不会遮住雕塑。

This is also a production decision.

这也是一个制作决定。

Exact text can remain editable, while the image carries the material and atmosphere.

精确文字可以保持可编辑，而图像承载材质和氛围。

If the title feels too dominant, change size, placement, or contrast and inspect the whole composition again.

如果标题太抢眼，就调整大小、位置或对比度，再检查整个构图。

Do not judge the text in isolation from the image.

不要脱离图像，单独评价文字。

The poster is the relationship between them, including the empty space that lets each element do its job.

海报是两者之间的关系，也包括让各元素发挥作用的空白。

Take a few seconds with both directions before I describe them.

在我描述之前，先花几秒看看两个方向。

Which would you put outside AFTER RAIN?

你会把哪一个放在《雨后》展厅外？

Choose privately first, so your answer is not just a response to mine.

先自己选择，避免答案只是对我的回应。

Then identify the visible feature that made you choose.

然后指出促使你选择的可见特征。

The photographic direction can invite attention to material and atmosphere.

摄影方向可以引导观众关注材质与氛围。

The collage direction can make fragility feel constructed through paper edges and layers.

拼贴方向可以通过纸边和层次，让脆弱感显得被建构出来。

Both can serve the brief, but they make different promises to a visitor.

两者都可能符合要求，但向观众许下的承诺不同。

A useful defense goes beyond realism versus abstraction.

有用的辩护，不应止于写实与抽象的区别。

Tell us what the audience is likely to notice or feel, and how the composition produces that reading.

请告诉我们，观众可能注意或感受到什么，以及构图如何产生这种解读。

We will use the disagreement to decide what to revise, rather than vote for a universally better image.

我们将用分歧来决定修改方向，而不是投票选出普遍更好的图像。

Selection is an artistic decision.

选择本身就是艺术决定。

Once generation gives us many plausible alternatives, choosing one determines what the work becomes.

当生成提供许多合理方案时，选中哪一个，就决定作品最终成为什么。

The reason should connect to the brief: perhaps the small object feels more exposed, or the torn edge makes fragility physically legible.

理由应与创作要求相连：可能是小物体更显得无所庇护，或撕裂边缘让脆弱感变得可触可见。

The rejected direction is useful evidence too.

被放弃的方向也是有用证据。

Explain what it does well and why you are choosing something else for this exhibition.

请解释它哪里做得好，以及为什么为这个展览选择了别的方案。

An unexpected model output can also change the direction, if you choose to develop it deliberately.

如果你决定有意识地发展一个意外输出，它也可以改变创作方向。

The key question is whether you can now articulate the intention and carry it through subsequent decisions.

关键是：你现在能否说清意图，并在后续决定中贯彻它？

Surprise can begin a work; it does not finish the artist's judgment.

惊喜可以开启作品，却不能替艺术家完成判断。

A useful critique names a visible feature, explains its effect, and proposes a revision.

有用的评论会指出可见特征、解释效果，并提出修改建议。

For example: the crisp reflection makes the collage feel photographic, so I would simplify its shape to match the flatter layers.

例如：清晰倒影让拼贴显得像摄影，所以我会简化它的形状，以配合更平面的层次。

Compare that with 'I do not like the reflection.' The first comment gives the artist a relationship to inspect and a possible next move.

与“我不喜欢倒影”相比，前一种评论给艺术家提供了可检查的关系，以及可能的下一步。

You can disagree about the intended effect, but make the disagreement specific.

你们可以对预期效果有不同看法，但要把分歧说具体。

This is a classroom critique framework rather than an official grading rubric.

这是课堂评论框架，不是正式评分标准。

Use it to connect evidence in the image to the idea the work is trying to communicate.

用它把图像中的证据，与作品试图传达的观念联系起来。

A curator asks, can the same sculpture feel more vulnerable without changing its shape?

策展人问：能否不改变形状，却让同一件雕塑显得更脆弱？

Lighting is one way to answer.

灯光是一种回答方式。

Our LightCtrl collaboration studies controllable relighting from a single image, including light direction, intensity, and color temperature.

我们合作开展的 LightCtrl 研究，探索从单张图像进行可控重光照，包括光线方向、强度和色温。

Follow the chair through the figure.

请沿着图中的椅子示例看。

The latent proxy encoder extracts compact physical cues.

潜在代理编码器提取紧凑的物理线索。

A lighting-aware mask guides the denoiser toward regions affected by the change, and preference optimization in the proxy branch supports physical consistency.

光照感知蒙版引导去噪器关注受变化影响的区域，代理分支中的偏好优化则支持物理一致性。

The method connects an interpretable lighting request with image generation.

这种方法把可解释的光照要求与图像生成连接起来。

For our pear, look for consequences rather than the word dramatic: where do highlights move, what happens to shadow, and how does the material read?

对于梨，不要只看“戏剧性”这个词，而要看后果：高光移向哪里、阴影如何变化、材质呈现如何？

Inferring geometry and material from one photograph is ambiguous, so the result still needs inspection.

从单张照片推断几何和材质存在歧义，因此仍需检查结果。

The exhibition example is our application of the idea, not an additional result reported by the paper.

展览示例是我们对这一思路的应用，不是论文另外报告的实验结果。

Choose between the two prepared directions for AFTER RAIN.

请在为《雨后》准备的两个方向中做出选择。

Which poster communicates fragility more clearly, and why?

哪张海报更清楚地传达脆弱感？为什么？

What does the rejected direction reveal about your choice?

被放弃的方向，揭示了你选择中的什么？

Which single revision would most change the audience’s reading?

哪一个单独的修改，会最大程度改变观众的解读？

Discuss these three questions with a specific visual feature in view.

请看着一个具体视觉特征，讨论这三个问题。

You can prefer quiet space or the instability of torn paper, but explain how that choice serves the exhibition.

你可以偏爱安静空间，也可以偏爱撕纸的不稳定感，但要解释这种选择如何服务于展览。

Pause the video here.

请在这里暂停视频。

Listen for a persuasive reason to choose the direction you initially rejected.

试着听取一个有说服力的理由，支持你最初拒绝的方向。

Consider what a process record lets us discuss.

想一想，过程记录能让我们讨论什么。

An artist can record the intention and invariants, the prompt and reference roles, and one observation about a rejected result.

艺术家可以记录意图与不变量、提示词与参考图角色，以及对一个被拒绝结果的观察。

That record explains what an attempt was testing.

这些记录说明了每次尝试在检验什么。

Without that context, a folder of attractive images can be difficult to interpret.

缺少背景信息，一文件夹好看的图像也可能很难解读。

We may not know which input changed or why a version was rejected.

我们可能不知道改了哪个输入，或为什么拒绝某个版本。

A few precise notes make a comparison more informative.

几条精确笔记，就能让比较更有信息量。

In a critique, this lets us ask about the artist’s decisions and the evidence behind them rather than guessing only from the final image.

在评论中，这让我们能够追问艺术家的决定及其证据，而不只是从最终图像猜测。

It also helps distinguish an intentional departure from an accidental change.

它也有助于区分有意偏离与意外变化。

Imagine we are standing between the two finished posters in a gallery review.

想象我们站在画廊评审现场，面前是两张完成的海报。

I would begin with a visible decision that carries the idea, then locate an unintended change, then propose one revision worth discussing.

我会先指出一个承载观念的可见决定，再找出一个意外变化，最后提出一个值得讨论的修改。

The order matters: the critique begins by understanding the work before prescribing a repair.

顺序很重要：评论应先理解作品，再给出修正方案。

We can hear two contrasting readings of the same poster.

对同一张海报，我们可能听到两种相反的解读。

One person may find the empty space quiet; another may find it emotionally distant.

一个人觉得留白安静，另一个人却觉得它情感疏离。

Ask which details support each reading.

请问，哪些细节支持各自的解读？

There is no image-making task here.

这里没有图像制作任务。

We are practicing how to make feedback useful to the next decision.

我们是在练习如何让反馈有助于下一次决定。

A good proposal names what would change and what effect we expect, so that a later version could confirm or challenge the reasoning.

好建议会说明改什么、预期产生什么效果，让后续版本能够证实或挑战这套推理。

Discuss these three questions, then pause the video for the gallery conversation.

请讨论这三个问题，然后暂停视频，进行画廊对话。

Before the artwork begins to move, defend a decision about the images.

在作品开始运动之前，请为一个图像决定提供理由。

How can a technically correct image fail as an artwork?

技术上正确的图像，为什么可能在艺术上失败？

Which visual rule should remain across a series?

一个系列中，哪条视觉规则应该保持？

What evidence makes a critique useful for the next revision?

什么证据能让评论对下一轮修改有用？

Discuss these three questions through one of the prepared posters.

请结合其中一张准备好的海报，讨论这三个问题。

Make words such as coherent or expressive concrete: locate the feature, describe its effect, and explain what changing it would do.

把“统一”“有表现力”等词说具体：指出特征、描述效果，并解释改变它会怎样。

Pause the video here.

请在这里暂停视频。

We will carry those artistic rules into the video section.

我们会把这些艺术规则带入视频部分。

Our image now has to survive time.

现在，图像必须经受时间的考验。

The subject can move, disappear, and return.

主体会运动、消失，然后返回。

We will examine selected frames from published research, follow the technical representations, and use a prepared AFTER RAIN scenario to decide what a convincing edit requires.

我们将查看已发表研究的精选帧、理解技术表示，并用准备好的《雨后》情境判断可信编辑需要什么。

Keep the preservation contract, but add an event that could break it.

保留编辑约定，再加入一个可能让它失效的事件。

A moving image creates new ways to break our contract.

动态图像会以新的方式破坏约定。

What new failures become possible when an image moves?

图像运动起来后，会出现哪些新的失败？

What should happen when the pear disappears behind a column?

梨消失在柱子后面时，应该发生什么？

Can four convincing frames prove that a video works?

四张可信的帧，能证明整段视频成立吗？

Discuss these three questions.

请讨论这三个问题。

Separate a correct change in visibility from an unwanted change in identity.

区分正确的可见性变化，与不希望发生的身份变化。

Name a moment you would need to see between the selected frames before accepting the shot.

接受这个镜头之前，请指出你需要在这些精选帧之间看到的某个时刻。

Pause the video here.

请在这里暂停视频。

Your prediction will give us a concrete test for the methods that follow.

你的预测会为后续方法提供一个具体测试。

A still image lets us choose a flattering instant.

静态图像让我们挑选一个最好看的瞬间。

A video makes the object keep its promises in the next frame.

视频则要求物体在下一帧继续信守承诺。

Look at the source sequence.

看原始序列。

As viewpoint and visibility change, we continue to recognize the subject.

随着视角和可见性变化，我们仍能认出主体。

An edit has to preserve that relationship while introducing its requested transformation.

编辑必须在引入所要求变换的同时，保留这种关系。

Remember our two clocks.

记住两种时钟。

The sampling steps describe how a model generates a result.

采样步描述模型如何生成结果。

The frames here describe time passing in the depicted scene.

这里的帧描述被描绘场景中时间的流逝。

A system may use many sampling steps to produce a short clip, but those steps are not extra seconds of action.

系统可能用很多采样步生成短片，但这些步数不是额外的剧情秒数。

Now the preservation contract must hold across an event, not just inside a frame.

现在，保留约定必须跨越一个事件成立，而不只是局限于单帧内部。

Would a video be perfectly consistent if every frame were identical?

如果每帧完全相同，视频就算完美一致吗？

Only if the intended scene were perfectly still.

只有目标场景本来就完全静止时才是。

In our moving shot, pose, viewpoint, illumination, and visibility should change.

在运动镜头中，姿态、视角、光照和可见性都应变化。

Consistency means that those changes remain coherent with the scene.

一致性意味着这些变化与场景保持协调。

The blue stem may become hidden during a turn.

蓝色柄可能在转动时被遮住。

That is different from its color changing without a lighting explanation.

这与没有光照原因却突然变色，是不同的。

A texture should move with its surface, rather than crawl independently across it.

纹理应该随表面一起移动，而不是独自在表面爬动。

If there is no convincing explanation, you may have found instability.

如果找不到可信解释，你可能发现了不稳定性。

This is why the goal cannot simply be to minimize change from one frame to the next.

因此，目标不能只是尽量减少相邻帧之间的变化。

Here is a prepared storyboard, not a generated video result.

这是准备好的分镜，不是生成视频的结果。

In the first view, the reflection suggests an object we have not fully seen.

第一个视角中，倒影暗示了一个我们尚未看全的物体。

A partial view then gives us enough evidence to make a guess.

接着，局部视角提供足够线索，让我们猜测。

The wide view finally changes our understanding of the setting.

最后的全景改变了我们对环境的理解。

The audience is doing something during those five seconds: forming an expectation, testing it, and revising it.

在这五秒中，观众一直在行动：建立预期、检验预期，然后修正它。

Compare this sequence with revealing everything in the first frame.

把这个序列与第一帧就揭示一切的版本比较一下。

The same sculpture could be present, but the experience would differ.

同一件雕塑可以都在场，但体验会不同。

For AFTER RAIN, the timing of information can carry fragility as strongly as material or lighting.

对于《雨后》，信息揭示的时机，可以像材质或灯光一样有力地承载脆弱感。

We should decide that structure before asking a tool to fill in motion.

要求工具补充运动之前，我们应该先决定这种结构。

Five seconds is short, but it can still have a structure.

五秒虽然短，却仍然可以有结构。

Our proposed first interval shows a reflection, giving the viewer clues.

我们设想的第一段展示倒影，给观众线索。

The second offers a partial view.

第二段提供局部视角。

The final interval reveals the wider setting and changes how the object is understood.

最后一段揭示更广阔的环境，改变人们对物体的理解。

Try another allocation and predict the effect.

试着重新分配时间，并预测效果。

A longer reflection might create uncertainty; a quick reveal might make the piece feel like a product shot.

更长的倒影段落可能制造不确定感；快速揭示则可能让它像产品广告镜头。

These timings are artistic proposals, not measured outputs from a video model.

这些时间安排是艺术提案，不是视频模型的实测输出。

By deciding the experience first, we can evaluate whether the generated motion and cuts support it.

先决定体验，才能评价生成的运动和剪辑是否支持它。

A technically smooth sequence can still have the wrong rhythm.

技术上流畅的序列，节奏仍可能不对。

The research begins with a surprisingly simple experiment: arrange video frames into a contact sheet and give that image to an image editor.

这项研究始于一个出人意料的简单实验：把视频帧排成接触印样，再交给图像编辑器。

This asks whether an existing image-editing capability can transfer across multiple views presented together.

它想检验：现有图像编辑能力，能否迁移到一起呈现的多个视角上？

Treat the result as evidence that motivates a research direction.

请把结果当作启发研究方向的证据。

A grid can make frames available in a shared context, but a successful-looking selection does not prove reliable video editing.

网格让多帧共享上下文，但一组看似成功的精选帧，并不能证明视频编辑可靠。

There may be flicker between the sampled frames.

采样帧之间可能仍有闪烁。

The next steps investigate how to represent video more effectively and adapt the model, rather than assuming a contact sheet alone solves time.

后续研究探索更有效的视频表示和模型适配，而不是假设接触印样本身就解决了时间问题。

But the missing intervals are where some of the most revealing failures live.

恰恰是在缺失的间隔中，可能藏着最有揭示性的失败。

A feature can jump, disappear, and return between the frames on this page.

某个特征可能在这一页的两帧之间跳动、消失，再返回。

Use the contact sheet for what it does well: compare appearance across selected moments and locate regions to inspect.

发挥接触印样的长处：比较选定时刻的外观，并定位需要检查的区域。

Then review full playback, with closer attention around turns and occlusion.

然后完整播放，特别关注转向和遮挡附近。

Think of a film review based only on publicity stills.

想象只凭宣传剧照来评论电影。

You might judge the costume and lighting, but you have not yet seen the performance.

你或许能评价服装和灯光，却还没有看过表演。

In generative video, that missing performance includes whether the same object continues to exist convincingly from moment to moment.

对生成视频而言，缺失的“表演”包括：同一个物体是否在每个时刻都可信地持续存在。

The grid can also be constructed in latent space.

网格也可以在潜在空间中构造。

Instead of only combining visible pixel images, the method encodes frames and arranges their tokens into a virtual image grid.

这种方法不只是组合可见像素图像，而是先编码各帧，再把 token 排列成虚拟图像网格。

The published example changes a bus into a graphics card.

论文中的示例把一辆公交车变成显卡。

The key idea is shared treatment of positions within that constructed representation.

核心思路是，在这种构造的表示中统一处理各位置。

Do not confuse this step with using a video VAE; they are different design decisions.

不要把这一步与使用视频 VAE 混为一谈；它们是不同的设计决定。

Putting frames near one another in a representation can help the model exchange information, but it does not impose a guarantee of physical continuity.

在表示中把帧放在一起，有助于模型交换信息，但不会强制保证物理连续性。

We still need to inspect what survives across changing views.

我们仍需检查视角变化时，哪些信息得以保留。

A virtual grid provides a way to arrange information for a model trained around image-like structure.

虚拟网格为围绕图像结构训练的模型，提供了一种信息排列方式。

It creates shared context and positional relationships among frame representations.

它在各帧表示之间建立共享上下文和位置关系。

That can be a useful bridge when reusing an image editor.

复用图像编辑器时，这可以成为有用的桥梁。

But arrangement alone does not enforce the laws of motion or the persistence of a hidden object.

但排列方式本身，不能强制执行运动规律或被遮挡物体的持续存在。

Those abilities depend on the model, adaptation, data, and other parts of the workflow.

这些能力取决于模型、适配、数据，以及工作流程中的其他部分。

Separate the representation choice from the capability claim.

请区分表示选择与能力主张。

A diagram can explain how frames enter a system without proving that the system handles every difficult event.

图表可以解释帧如何进入系统，却不能证明系统处理得了所有困难事件。

That proof would require appropriate outputs and evaluation.

证明这一点，需要适当的输出和评价。

This architecture asks whether an image editor's abilities can be reused for video.

这个架构检验图像编辑器的能力能否复用于视频。

The video VAE encodes a clip into a compressed latent representation.

视频 VAE 将片段编码成压缩的潜在表示。

Learned projections connect those video latents with the image-editing transformer, so the model can operate on a compatible arrangement of information.

学习得到的投影把视频潜变量与图像编辑 Transformer 连接起来，让模型处理兼容的信息排列。

Trace that main route before reading the smaller branches.

阅读较小分支之前，先沿着这条主路径看。

Adaptation connects them; the paper explores LoRA and full training choices, with an optional enhancement stage.

适配把它们连接起来；论文探索了 LoRA 和全量训练，并包含可选的增强阶段。

The attractive idea is reuse of editing knowledge.

吸引人的想法是复用编辑知识。

The test is whether that reuse preserves the temporal relationships our shot needs.

真正的检验是：这种复用是否保留了镜头需要的时间关系？

Architecture explains where information flows; the edited clip shows whether the intended event survives.

架构解释信息流向哪里；编辑后的片段展示预期事件是否得以保留。

The image model and video VAE do not necessarily speak the same representational language.

图像模型与视频 VAE 未必使用相同的表示语言。

Learned projections help connect them.

学习得到的投影帮助连接二者。

Adaptation then allows the editing transformer to operate usefully with the video representation.

之后，适配让编辑 Transformer 能有效使用视频表示。

This is a common engineering pattern: reuse a capable component while learning the interface and behavior needed for another task.

这是一种常见工程模式：复用有能力的组件，同时学习另一项任务所需的接口和行为。

The benefit is not automatic.

收益并不会自动出现。

A projection must preserve useful information, and the adapted model must learn how the new structure relates to edits.

投影必须保留有用信息，适配后的模型也必须学会新结构如何与编辑相关联。

In the published pipeline, decoding returns the edited representation to video.

在已发表的流程中，解码将编辑后的表示还原为视频。

Follow the information through every stage rather than treating the name of the reused model as an explanation by itself.

请追踪信息经过每个阶段，而不要把被复用模型的名字本身当作解释。

Here is a counting puzzle with a small trap.

这里有一个带小陷阱的计数题。

In this example, forty-five pixel frames are compressed with a special first frame and a factor of four for the remaining temporal groups.

在这个例子中，四十五个像素帧采用特殊首帧处理，其余时间组按四倍压缩。

Forty-five is four times eleven, plus one.

四十五等于四乘十一，再加一。

The latent sequence therefore has eleven plus one, or twelve frames.

因此，潜在序列有十一加一，即十二帧。

Those twelve latent frames can be arranged in the illustrated three-by-four virtual grid.

这十二个潜在帧可以排成图中的三乘四虚拟网格。

Simply dividing forty-five by four would miss the first-frame convention.

简单地用四十五除以四，会忽略首帧约定。

This arithmetic belongs to the representation used in this example; it is not a rule for every video model.

这个算法属于本例采用的表示，并不是所有视频模型的通用规则。

A small detail in temporal compression changes what the editing model actually receives.

时间压缩中的一个小细节，会改变编辑模型实际收到什么。

Solve the frame-count relationship step by step.

一步步解出帧数关系。

Start with four k plus one equals forty-five.

从 4k 加一等于四十五开始。

Subtract one to obtain forty-four, then divide by four to get k equals eleven.

减一得到四十四，再除以四，得到 k 等于十一。

The latent sequence contains k plus one frames, so its length is twelve.

潜在序列包含 k 加一帧，因此长度为十二。

This arithmetic encodes the special handling of the first frame in the stated convention.

这个计算反映了该约定对首帧的特殊处理。

It is different from simply dividing the total number of pixel frames by four.

它不同于直接把像素帧总数除以四。

Understanding them can prevent mistakes when arranging latent frames into a virtual grid or comparing the representation with the original clip.

理解这些关系，可以避免在排列潜在帧网格、或与原始片段比较时出错。

The requested edit turns the scene into a cyberpunk workshop with holographic documents.

编辑要求是把场景变成赛博朋克工作室，并加入全息文档。

Compare the results before focusing on the setting labels.

先比较结果，再关注设置标签。

Which version carries out the requested transformation more completely, and what visible evidence supports your answer?

哪个版本更完整地实现了要求的变换？有什么可见证据？

The paper presents this case as an example where classifier-free guidance improves edit completeness.

论文用这个案例说明，无分类器引导可以改善编辑完整性。

Its baseline retains source latents, and its implementation includes rescaling.

其基线保留了原始潜变量，实现中还包含重缩放。

This does not establish one universally best guidance value.

这并不意味着存在一个普遍最优的引导值。

Guidance also has a computation cost when it requires another model pass.

如果引导需要额外一次模型前向计算，也会带来计算成本。

Relate the example back to our equation: changing a prediction combination affects how strongly the result follows the condition.

联系之前的方程：改变预测组合，会影响结果遵循条件的力度。

An ablation asks what changes when one component is removed.

消融实验追问：移除某个组件后，会发生什么变化？

In the displayed example, guidance makes more of the requested transformation visible.

在展示的例子中，引导让更多要求的变换可见。

That is useful evidence about this comparison.

这是关于这次比较的有用证据。

It is not yet a measurement of reliability across the kinds of shots we might produce for an exhibition.

但它还不是对我们可能用于展览的各类镜头的可靠性测量。

As a reader, separate the observation from the next question.

作为读者，请区分观察结果与下一步问题。

We can observe a more complete edit here.

我们可以观察到，这里的编辑更加完整。

We would still want to know what happens across other clips, how often preservation suffers, and what the computation costs.

但仍想知道：其他片段会怎样、保留效果多常受损，以及计算成本是多少。

You can learn a mechanism from a selected example while designing a stronger test for the decision you actually need to make.

你可以从精选示例中学习机制，同时为真正需要做出的决定设计更强的测试。

This is a local edit: the sheep's face changes and receives a white star-shaped patch.

这是一次局部编辑：绵羊的脸发生变化，增加了一块白色星形斑纹。

Look at the patch across poses.

观察不同姿态下的这块斑纹。

Does it remain attached to the same facial region, with a plausible change in apparent shape as the head turns?

它是否始终附着在同一面部区域，并在头部转动时合理改变表观形状？

The most attractive frame is not enough.

仅看最漂亮的一帧是不够的。

An identity marker can slide, disappear, or change shape at another moment.

身份标记可能在另一个时刻滑动、消失或变形。

Selected frames let us ask the right questions, but the full clip is needed to check continuity between them.

精选帧让我们提出正确问题，但要检查帧间连续性，仍需完整视频。

For our discussion, identify a small recognizable feature that will make identity drift easier to notice.

请为讨论找一个可辨认的小特征，让身份漂移更容易被发现。

Choose a distinctive feature and follow it through the shot: its attachment to the subject, its shape, and its reappearance after partial visibility.

选一个独特特征，并在镜头中追踪它：与主体的附着关系、形状，以及部分遮挡后的再次出现。

For our pear, the blue stem is a useful witness.

对这只梨而言，蓝色柄是有用的见证者。

It may become hidden as the camera moves; we should allow that.

随着相机移动，它可能被遮住；这是应当允许的。

When it returns, it should still belong to the same object.

当它返回时，仍应属于同一个物体。

A marker that slides across the surface or reappears in a new shape tells us something a general impression of smooth motion might miss.

如果标记在表面滑动，或以新形状返回，就揭示了整体流畅感可能掩盖的问题。

We are using the marker as a diagnostic aid, while still judging the whole object's identity and the shot's intended motion.

我们用标记辅助诊断，同时仍评价整个物体的身份和镜头预期运动。

Here the instruction changes the whole sequence into a minimal monochrome sketch.

这里的指令把整个序列变成极简单色素描。

Local object replacement and global stylization allow different degrees of visual freedom.

局部物体替换与整体风格化，允许的视觉自由程度不同。

Even with a large style change, motion and scene structure should remain readable.

即使风格变化很大，运动和场景结构也应保持可读。

Look for rules across frames: line density, silhouette treatment, and the handling of depth.

寻找跨帧规则：线条密度、轮廓处理，以及深度的表现方式。

Do they feel like one visual language?

它们是否像同一种视觉语言？

This connects directly to our collage discussion.

这与我们的拼贴讨论直接相连。

A style is more useful to an art director when described through operations that can persist through time.

当风格被描述为能够持续存在于时间中的操作时，它对艺术总监更有用。

If every frame reinvents those operations, the result may feel unstable even when each still is appealing.

如果每一帧都重新发明这些操作，即使每张静帧好看，整体也可能不稳定。

A video can preserve object motion while its style flickers.

视频可以保留物体运动，风格却仍然闪烁。

Line density may jump, a paper texture may crawl, or shading may switch between flat and volumetric treatments without a scene explanation.

线条密度可能跳变，纸张纹理可能爬动，明暗处理也可能无故在平面与立体之间切换。

Review style as a set of temporal rules, just as we reviewed it across surfaces in the collage.

像审查拼贴不同表面的风格那样，把视频风格作为一组时间规则来审查。

Some variation is appropriate when the camera or light changes.

相机或光线变化时，某些变化是合理的。

The question is whether the material language remains coherent.

问题是：材质语言是否仍然连贯？

If your artwork depends on a particular drawing or collage treatment, this check is as important as whether the subject stays in the same place.

如果作品依赖特定绘画或拼贴处理，这项检查与主体位置是否稳定同样重要。

Motion correctness alone does not establish visual consistency.

运动正确，本身并不能证明视觉一致。

If we subtract one frame from the next, a perfectly correct camera movement can create a large difference.

如果将相邻帧相减，完全正确的相机运动也会产生很大差异。

The same surface has simply moved to another pixel location.

同一个表面，只是移动到了另一个像素位置。

A more meaningful comparison first aligns corresponding visible content, then measures the remaining appearance difference.

更有意义的比较，是先对齐对应的可见内容，再测量剩余外观差异。

The teaching equation uses a motion warp for that alignment and a visibility mask to exclude occluded regions.

这个教学公式使用运动变换来对齐，并用可见性蒙版排除被遮挡区域。

We should not demand agreement for content that is hidden or newly revealed.

对于隐藏或刚被揭示的内容，不应要求直接一致。

This is an illustrative metric, not a claim about the exact evaluation used by the paper.

这是示意性指标，不是在声称论文使用了这一精确评测。

A low difference can also reward a video that barely moves.

低差异也可能奖励几乎不动的视频。

A metric must be checked against the task it is supposed to represent.

必须对照指标所要代表的任务，检查指标本身。

Imagine two candidates for our five-second shot.

想象五秒镜头的两个候选版本。

One shows a coherent camera move around the pear with modest frame differences.

一个展示相机连贯地绕梨移动，帧间差异适中。

The other repeats a single beautiful frame.

另一个反复显示同一张漂亮图像。

A naive difference score might prefer the frozen version.

简单的差异分数，可能偏爱冻结版本。

Our audience would immediately notice that the intended reveal never happened.

观众却会立刻发现，预期的揭示根本没有发生。

That is a useful example of optimizing the measurement while missing the purpose.

这是优化测量指标却错失目的的一个例子。

We need evidence for both coherent appearance and requested motion.

我们需要外观连贯和所要求运动两方面的证据。

Neither can substitute for the other.

两者不能相互替代。

Each time, one easy-to-observe quality tempted us to stand in for the whole brief.

每一次，我们都容易让某个易于观察的品质，代表整份创作要求。

Good evaluation keeps the intended result in view, especially when a convenient score seems reassuring.

良好的评测始终关注预期结果，尤其当方便的分数令人安心时。

An edited keyframe can make an art direction easier to approve.

编辑后的关键帧，可以让艺术方向更容易获得认可。

Instead of describing every material and lighting choice in words, we can point to an image and say, this is the appearance we want.

不必用文字描述每个材质和灯光决定，我们可以指着一张图说：这就是想要的外观。

The workflow then carries that approved look through the source clip.

工作流程再把获批的外观延续到原始片段中。

Follow the four stages: select a representative frame, edit and inspect its appearance, propagate the look, and review the resulting motion.

请跟随四个阶段：选代表帧，编辑并检查外观，传播外观，再审查运动结果。

Will the material persist during a turn?

转向时，材质能保持吗？

Will the identity survive a column passing in front of it?

柱子从前方经过时，身份能保留吗？

This distinction is central to the Runway reading later: an appealing interaction pattern still needs a shot review suited to the artwork.

这个区别是后面 Runway 阅读的核心：吸引人的交互方式，仍需要适合这件作品的镜头审查。

What if the approved image is so influential that the requested event never happens?

如果获批图像影响力太强，以至于要求的事件根本没发生，怎么办？

Our AlignVid collaboration studies that tension in image-to-video generation.

我们合作开展的 AlignVid 研究，探索图生视频中的这种张力。

In the published examples, a baseline omits a sunflower or leaves a person standing; the corresponding AlignVid results implement more of the requested event.

论文示例中，基线漏掉了向日葵，或让人物一直站着；对应的 AlignVid 结果实现了更多要求的事件。

The intervention scales queries or keys in selected attention blocks and denoising steps without retraining.

这种干预无需重新训练，只在选定注意力模块和去噪步骤中，缩放查询或键。

In a scalar form, scaling Q by gamma changes the weights to softmax of gamma times Q K transpose over square root d.

在标量缩放形式下，用 gamma 缩放 Q 后，权重变成 softmax(gamma × QKᵀ / √d)。

It changes attention concentration.

它改变了注意力的集中程度。

Classifier-free guidance instead combines predictions with different conditioning; these are distinct operations.

无分类器引导则组合不同条件下的预测；两者是不同操作。

For an artist, the interesting failure is a faithful-looking image that refuses to do what the scene requires.

对艺术家来说，有意思的失败是：图像看起来很忠实，却拒绝做场景要求的事。

The published frames illustrate that tension; a finished shot still needs temporal review.

已发表的帧展示了这种张力；完成的镜头仍需进行时间连续性审查。

Our five-second brief changes the pear's body to ivory ceramic while retaining the blue stem, camera movement, and position.

我们的五秒要求，是把梨的主体变为象牙白陶瓷，同时保留蓝色柄、相机运动和位置。

Reflections may adapt to the new material.

倒影可以适应新材质。

The pear must remain the same object after passing behind a column.

梨经过柱子后面再出现时，必须仍是同一个物体。

Which clause is hardest to verify?

哪一条最难验证？

The reappearance is a strong candidate, because the system must maintain identity through a period of invisibility.

再次出现很可能最难，因为系统必须在一段不可见期间保持身份。

This is more demanding than transferring a visible color from frame to frame.

这比逐帧传递可见颜色更具挑战。

Begin with one clear transformation in a short clip.

先在短片中做一个明确的变换。

Then make the critical visibility event part of the review, rather than discovering it only after choosing a favorite result.

然后把关键可见性事件纳入审查，而不是选好最喜欢的结果后才发现它。

The column is the moment of truth.

柱子就是关键考验。

Before the pear disappears, we can inspect its silhouette, material, and stem.

梨消失前，我们能检查它的轮廓、材质和柄。

During occlusion, there may be nothing visible to compare.

遮挡期间，可能没有可见内容可供比较。

When it returns, the model has to make the same object convincing again.

返回时，模型必须再次让同一个物体显得可信。

Do not reject correct invisibility as a failure.

不要把正确的不可见状态当作失败。

Instead, compare the identity before and after the event, including how the object emerges at the boundary.

而应比较事件前后的身份，包括物体如何从边界处显露。

A smooth-looking clip can still reveal a subtly redesigned pear on the other side.

看起来流畅的视频，也可能在另一侧露出一只被悄悄重新设计的梨。

We are not merely watching for anything strange; we are testing the preservation requirement at the point where the shot makes it hardest to satisfy.

我们不只是寻找怪异之处，而是在镜头最难满足要求的地方，检验保留约束。

Adding a small object creates several new relationships.

添加一个小物体，会创造多种新关系。

In this published example, the edit adds a drone.

在这个已发表示例中，编辑添加了一架无人机。

Even if the drone looks convincing by itself, its scale, placement, and movement must fit the scene.

即使无人机本身很可信，它的尺度、位置和运动也必须适合场景。

Imagine a camera moving forward while the added object changes size in the wrong direction.

想象相机向前移动，新增物体的大小却朝错误方向变化。

The object might be beautifully rendered, but its relationship with the camera would expose the edit.

物体可能渲染得很漂亮，但它与相机的关系会暴露编辑。

Or it might hover at an unintended location relative to other objects.

或者，它可能相对其他物体悬停在错误位置。

When you review an addition, evaluate those relationships across time.

审查新增物体时，要跨时间评价这些关系。

A close-up of the inserted object cannot answer every question about whether it belongs.

仅看新增物体的特写，无法回答它是否融入场景的所有问题。

An added object must fit the camera as well as the scene.

新增物体既要符合场景，也要符合相机。

A convincing texture and shape in one still do not establish a coherent trajectory.

单张静帧中可信的纹理和形状，不能证明运动轨迹连贯。

If the camera approaches, the apparent size and position of the addition should evolve in a compatible way.

如果相机靠近，新增物体的表观大小和位置应相应变化。

The exact expectation depends on whether the object is stationary or moving independently.

具体预期取决于物体是静止，还是独立运动。

That is why the brief should specify the intended relationship.

因此，创作说明应该明确预期关系。

Inspect the addition relative to nearby objects and the background, not only in a crop.

要相对附近物体和背景检查新增内容，而不只看裁剪图。

A production-quality edit is a collection of relationships that survive over time, rather than a new object that looks impressive in isolation.

制作级编辑是一组经得起时间考验的关系，而不是一个单独看很惊艳的新物体。

Review the whole shot at normal speed, then inspect moments where failure is most likely.

先以正常速度看完整个镜头，再检查最可能失败的时刻。

Pay attention to identity, intended motion, occlusion, boundaries, and the ending.

关注身份、预期运动、遮挡、边界和结尾。

High-motion content is a limitation the research authors specifically identify.

高运动量内容，是研究作者明确指出的局限。

If a clip fails, choose a revision based on the failure.

片段失败时，要根据具体失败选择修改。

You might reduce the transformation, shorten the segment, add a reference frame where supported, or repair part of the result through conventional compositing.

可以减小变换幅度、缩短片段、在支持时添加参考帧，或用传统合成修复部分结果。

Another generation is useful when it tests a reasoned hypothesis.

当新一轮生成用于检验有理由的假设时，它才有用。

Do not let repeated attempts distract you from whether the piece still communicates the artistic intention you began with.

不要让反复尝试分散注意力，忘记作品是否仍在传达最初的艺术意图。

A failed shot is information about the workflow.

失败镜头提供了关于工作流程的信息。

If the appearance is wrong from the beginning, revising the keyframe or the material direction may help.

如果一开始外观就不对，修改关键帧或材质方向可能有帮助。

If the first frame is convincing but identity breaks after occlusion, the problem calls for temporal review and a more suitable propagation or editing approach.

如果首帧可信，但遮挡后身份破坏，就需要时间连续性审查，以及更合适的传播或编辑方法。

If a region must remain exact, direct compositing may be part of the solution.

如果某个区域必须完全不变，直接合成可以是解决方案的一部分。

We can also split a complicated transformation into stages, provided the transitions remain coherent.

也可以把复杂变换拆成多个阶段，但必须保证过渡连贯。

The important move is to connect the observed failure to the next intervention.

关键是把观察到的失败，与下一步干预联系起来。

Repeating a vague request with more enthusiasm does not use what the failed shot has taught us.

更热情地重复模糊要求，并没有利用失败镜头教给我们的东西。

Return to the proposed five-second ceramic-pear shot.

回到拟议的五秒陶瓷梨镜头。

Which moment would you inspect most closely?

你会最仔细检查哪个时刻？

What should remain recognizable after the column?

经过柱子后，什么应该仍然可辨认？

What failure would make you reject an attractive clip?

什么失败会让你拒绝一段好看的视频？

Discuss these three questions by describing an event and the evidence you would seek.

请通过描述事件和所需证据，讨论这三个问题。

If your answer is temporal consistency, make it visible: what changes, when, and why would that violate the intended shot?

如果你的答案是“时间一致性”，请让它可见：什么在何时改变，为什么违反了镜头意图？

Pause the video here.

请在这里暂停视频。

Compare your acceptance criteria with another person’s before we examine the compact review plan.

在查看简要审查方案之前，先与另一位同学比较验收标准。

Here is a compact test plan.

这是一个简要测试方案。

The request is an ivory ceramic body with the blue stem retained.

要求是象牙白陶瓷主体，并保留蓝色柄。

Camera motion and object motion should follow the source.

相机运动和物体运动应遵循原片。

The clip includes a column so that we can inspect a difficult reappearance.

片段中包含柱子，让我们检查困难的再次出现过程。

The evidence includes normal-speed playback and a closer look before and after that event.

证据包括正常速度播放，以及事件前后的仔细观察。

We also inspect reflections because a material change has optical consequences.

还要检查倒影，因为材质变化会产生光学后果。

This plan does not require a long benchmark suite.

这个方案不需要庞大的基准测试集。

It simply makes the request, the likely failure, and the acceptance evidence explicit.

它只是明确要求、可能的失败，以及验收证据。

You can use the same structure to test a different subject or visual transformation.

你可以用同样的结构测试其他主体或视觉变换。

Let us check whether the mechanisms are now useful without a model in front of us.

现在没有模型在面前，让我们检验这些机制是否真的有用了。

Consider the three questions on screen.

请考虑屏幕上的三个问题。

A material change reaches the reflection: explain why.

材质变化会影响倒影：请解释原因。

A subject reference and LoRA both help represent a concept: explain the difference.

主体参考图和 LoRA 都有助于表达概念：请解释两者区别。

Four frames look excellent: explain why the clip can still fail.

四帧看起来很出色：请解释整段视频为什么仍可能失败。

Take a short silent beat before answering.

回答前，先安静思考片刻。

The next slide names the decisions hiding inside the questions.

下一页会指出这些问题中隐藏的决定。

If you get stuck, locate the type of problem first: the physical scene, the model's information, or evidence across time.

如果卡住了，先定位问题类型：物理场景、模型获得的信息，还是跨时间的证据。

That is often enough to begin a clear answer.

通常，做到这一步就足以开始给出清晰答案。

Here are the decisions those questions conceal.

这些问题隐藏着以下决定。

If the reflection must change with the new material, we need to define the permitted region and consequences of the edit.

如果倒影必须随新材质变化，就需要定义编辑允许的区域和连带后果。

If we want a concept for this request, a reference can condition generation; if we want learned adaptation, LoRA changes selected weights.

如果只为这次要求表达概念，参考图可提供生成条件；如果需要学习式适配，LoRA 会改变选定权重。

If continuity matters, the evidence must include the intervals between our chosen stills.

如果重视连续性，证据就必须包含精选静帧之间的间隔。

Each decision rules out a tempting shortcut.

每个决定，都排除了一条诱人的捷径。

Freezing every surrounding pixel may freeze the wrong reflection.

冻结周围每一个像素，可能冻结错误的倒影。

Calling every supplied image training confuses conditioning with adaptation.

把所有输入图像都称为训练，会混淆条件控制与适配。

Calling four stills a successful video omits time.

把四张静帧称为成功视频，就遗漏了时间。

Use this slide to repair your explanation, rather than memorize a preferred sentence.

请用这一页修正你的解释，而不是背诵某句标准答案。

The material changes how light interacts with the pear, so its visible consequences can extend into the reflection.

材质改变光与梨的相互作用，因此可见后果会延伸到倒影。

A reference conditions an output; LoRA learns a low-rank update to selected model weights.

参考图为输出提供条件；LoRA 为选定模型权重学习低秩更新。

Four good frames leave unobserved transitions, including events such as occlusion where identity can fail.

四张好帧仍留下未观察的过渡，包括可能导致身份失败的遮挡事件。

Those are compact answers, but each has an application.

这些答案很简短，但每个都有实际用途。

They help us write a better preservation contract, choose between two kinds of intervention, and design a more revealing review.

它们帮助我们写出更好的保留约定、在两类干预中选择，并设计更有揭示性的审查。

If your answer used different words and preserved those distinctions, it works.

如果你用了不同措辞，但保留了这些区别，答案就成立。

Tomorrow's interface may rename its controls, but the difference between specifying a request, adapting a model, and checking its result will still matter.

明天的界面可能给控件改名，但提出要求、适配模型与检查结果之间的区别，仍然重要。

We can check an edit at three levels.

我们可以在三个层面检查编辑。

First, did the requested transformation occur?

第一，所要求的变换发生了吗？

Second, did the necessary invariants survive?

第二，必要的不变量保留了吗？

Third, does the result serve the artwork's intention?

第三，结果是否服务于作品意图？

A result can pass the first level and fail the second, as when the correct material appears on a different object.

结果可能通过第一层却不通过第二层，例如正确材质出现在了另一个物体上。

It can pass both technical levels and still fail the artistic one, as when the chosen effect undermines the intended mood.

也可能通过两个技术层面，却在艺术层面失败，例如所选效果削弱了预期情绪。

Keeping the levels separate makes critique more precise.

区分这些层面，会让评论更精确。

It also explains why neither a generic aesthetic score nor a strict pixel comparison can answer every question we care about.

这也解释了为什么通用审美分数和严格像素比较，都无法回答我们关心的所有问题。

We have spent the lecture making the editing request more precise.

整堂课，我们都在让编辑要求更精确。

Now consider who gets to choose the request in the first place.

现在想想：一开始由谁来选择提出什么要求？

A system can offer convincing alternatives, but somebody still decides which intention matters and which evidence counts as success.

系统可以提供可信方案，但仍需要有人决定什么意图重要，以及什么证据才算成功。

The three readings let us examine that responsibility from different positions.

三篇阅读让我们从不同位置审视这种责任。

Pachocki raises questions about capable AI systems and human values.

Pachocki 提出了能力强大的 AI 系统与人类价值观之间的问题。

Koe asks readers to reconsider their own goals and habits.

Koe 请读者重新审视自己的目标与习惯。

Runway presents a workflow for turning an approved image into a video edit.

Runway 展示把获批图像转化为视频编辑的工作流程。

Keep our pear in mind as we read them: we can delegate parts of its production while still arguing about what the artwork should become.

阅读时请记住这只梨：我们可以委托部分制作，同时仍讨论作品应该成为什么。

We have an approved appearance and a moving shot to judge.

现在，我们有了获批外观，以及需要评价的运动镜头。

What can an approved keyframe establish, and what can it not?

获批关键帧能确定什么，又不能确定什么？

Why might a low frame-difference score reward a bad video?

为什么低帧差分数可能奖励一段坏视频？

Which evidence would convince you that identity survived motion?

什么证据能说服你，身份经受住了运动？

Discuss these three questions.

请讨论这三个问题。

Your answer should account for the requested event as well as the subject’s appearance.

答案既要考虑主体外观，也要考虑所要求的事件。

Use the column, a turn, or the sketch treatment as a concrete example.

可以用柱子、转向或素描处理作为具体例子。

Pause the video here.

请在这里暂停视频。

We will next ask how the readings change our view of responsibility for those decisions.

接下来，我们将追问阅读如何改变我们对这些决定所涉及责任的看法。

These first two readings ask about direction from different sides.

前两篇阅读从不同方面追问方向。

An Alien Mind considers the relationship between an AI system achieving a goal and respecting human values.

《An Alien Mind》思考 AI 系统实现目标与尊重人类价值观之间的关系。

Dan Koe's essay invites people to examine the goals and habits directing their own lives.

Dan Koe 的文章邀请人们审视引导自己生活的目标和习惯。

We will use its proposal as material for critical discussion of creative practice.

我们把他的主张作为批判性讨论创作实践的材料。

Find a specific claim, explain your interpretation, and connect it to a decision from the lecture.

找出一个具体主张，解释你的理解，并联系课堂中的一个决定。

A fictional artist is fine for discussing personal direction.

讨论个人方向时，可以使用虚构艺术家的例子。

The fifth quiz question asks you to reflect on one reading of your choice.

测验第五题要求你从自选的一篇阅读出发进行反思。

Our class discussion can compare all three perspectives without requiring everyone to disclose personal experiences.

课堂可以比较三种视角，而不要求每个人披露个人经历。

The third reading moves from questions about direction to a concrete editing workflow.

第三篇阅读从方向问题，转向具体编辑流程。

Runway's announcement of Aleph 2.0 and Edit Studio describes establishing an appearance in an edited image and applying that change through video.

Runway 关于 Aleph 2.0 和 Edit Studio 的公告，描述了先在编辑图像中确定外观，再把变化应用到视频的方法。

It connects directly to our keyframe discussion.

它与关键帧讨论直接相连。

Read it with the ceramic pear and column in mind.

阅读时，请想着陶瓷梨和柱子。

Which decision does the interface make easier to express?

这个界面让哪项决定更容易表达？

Which result would you still need to watch before approval?

批准之前，哪些结果仍需要亲眼播放检查？

The page is a vendor's account, so its demonstrations illustrate claimed capabilities rather than supply an independent comparative evaluation.

这是一篇厂商介绍，因此演示说明的是其声称的能力，而不是独立的比较评测。

We can learn from the interaction design and still propose an occlusion test that asks whether the workflow meets this exhibition's particular needs.

我们可以借鉴交互设计，同时提出遮挡测试，检验流程是否满足这个展览的具体需要。

These technical papers are optional companions to the discussion readings.

这些技术论文，是讨论阅读之外的可选配套材料。

Choose according to the question you want to investigate.

请根据你想研究的问题选择。

InstructPix2Pix helps explain edit supervision; Flow Matching develops the generative training framework.

InstructPix2Pix 帮助解释编辑监督；Flow Matching 发展了生成训练框架。

Qwen-Image-2.0 and Qwen-Video-Edit provide recent architecture examples.

Qwen-Image-2.0 和 Qwen-Video-Edit 提供了近期架构示例。

You do not need to read all of them before trying the class discussion.

参加课堂讨论之前，不必读完全部论文。

Start with a question, locate the part of the paper that addresses it, and distinguish the proposed method from the authors' evidence.

从一个问题开始，找到论文中回应它的部分，并区分提出的方法与作者的证据。

The slide notes retain the other references on multi-image inputs, rewards, composition, and video editing.

幻灯片备注保留了多图输入、奖励、构图和视频编辑方面的其他参考资料。

The three readings provide different kinds of material.

三篇阅读提供不同类型的材料。

A research leader's perspective develops an argument about alignment and future development.

研究领导者的观点文章，提出关于对齐和未来发展的论证。

A reflective essay offers a way to examine personal direction.

反思性文章提供一种审视个人方向的方法。

A vendor announcement presents a workflow and promotes its capabilities.

厂商公告展示工作流程，并推广其能力。

Read each according to what it can support.

阅读每种材料时，要考虑它能够支持什么。

Identify forecasts, personal claims, and demonstrations rather than treating every sentence as the same kind of evidence.

识别预测、个人主张和演示，不要把每句话都当作同类证据。

You may disagree with an author and still find a useful question.

你可以不同意作者，却仍从中找到有用的问题。

Our synthesis is practical: what do you want to make, what will you delegate, and what evidence will you use to decide whether the collaboration served that intention?

我们的综合问题很实际：你想创作什么、委托什么，以及用什么证据判断合作是否服务于这一意图？

This technical extension separates image guidance and text guidance in InstructPix2Pix.

这个技术延伸，区分 InstructPix2Pix 中的图像引导和文本引导。

Begin with the prediction using neither condition.

从不使用任何条件的预测开始。

Add a scaled difference for including the image, then another scaled difference for adding the text alongside that image.

加上引入图像后的缩放差值，再加上在图像之外引入文本后的另一个缩放差值。

Set both scales to one and follow the cancellation: the intermediate terms disappear, leaving the fully conditioned noise prediction.

把两个强度都设为一，观察抵消：中间项消失，只留下完全条件化的噪声预测。

This is a useful check that you understand the expression.

这是检查你是否理解表达式的有用方法。

Epsilon here denotes a noise prediction, unlike the velocity in our flow-matching slides.

这里的 epsilon 表示噪声预测，与流匹配页面中的速度不同。

These controls belong to this formulation and should not be assumed to map directly onto every current editor's interface.

这些控制属于这一公式，不应假设它们直接对应所有当前编辑器的界面。

Now set both guidance scales to one and follow the cancellation.

现在把两个引导强度都设为一，观察抵消。

The initial no-condition prediction cancels its negative copy in the first difference.

初始无条件预测，与第一个差值中的负项抵消。

The image-only prediction then cancels its negative copy in the second difference.

随后，仅图像条件的预测，与第二个差值中的负项抵消。

The result is the prediction with both image and text conditions.

最终得到同时使用图像和文本条件的预测。

That gives us a reference point for understanding the two controls.

这为理解两个控制提供了参考点。

If the text scale were zero while the image scale stayed one, we would instead recover the image-only prediction.

如果文本强度为零、图像强度仍为一，就会得到仅使用图像条件的预测。

It shows what information each difference adds.

这说明了每个差值添加什么信息。

Keep epsilon's meaning explicit here: this InstructPix2Pix expression predicts noise, whereas the earlier flow-matching example predicted velocity.

请明确 epsilon 的含义：这里的 InstructPix2Pix 公式预测噪声，而前面的流匹配例子预测速度。

Read the pseudocode in two groups.

把伪代码分成两组来读。

First we prepare an example: source, instruction, target, conditions, and target latent.

首先准备样本：原图、指令、目标、条件，以及目标潜变量。

Then we sample noise and time, construct the intermediate latent, predict velocity, and compare that prediction with the known training target.

然后采样噪声和时间，构造中间潜变量，预测速度，并将预测与已知训练目标比较。

The final update changes trainable weights based on the loss.

最后根据损失更新可训练权重。

At inference, there is no known edited target to supply in this way.

推理时，并没有已知编辑目标可以这样提供。

We instead follow the learned generative predictions from a starting state.

我们从初始状态出发，遵循学习得到的生成预测。

This pseudocode explains the conceptual loop; it omits practical engineering such as batches, precision, and scheduling.

这段伪代码解释概念循环，省略了批处理、精度和调度等实际工程细节。

Its most important distinction is what information is available during learning versus use.

最重要的区别，是学习与使用时分别有哪些信息可用。

Look for the line that disappeared.

请找出消失的那一行。

The inference loop has no known edited target and no loss that updates weights.

推理循环没有已知编辑目标，也没有用于更新权重的损失。

Instead, we encode the source and instruction, begin with noise, repeatedly predict a velocity and move the latent, then decode the result.

我们编码原图和指令，从噪声开始，反复预测速度并移动潜变量，最后解码结果。

Training uses examples to adjust the learned model.

训练用样本调整学习到的模型。

In this simplified inference loop, we use that model to construct an output we do not yet possess.

在这个简化推理循环中，我们用该模型构造尚未拥有的输出。

The code is conceptual; practical systems use their own representations and sampling schedules.

代码是概念性的；实际系统使用各自的表示和采样调度。

But you can now explain why providing another reference changes the information available for a request without automatically becoming a model-training operation.

但你现在可以解释：增加参考图会改变请求可用的信息，却不会自动变成模型训练。

It is the same distinction we used when choosing between references and LoRA.

这与选择参考图还是 LoRA 时使用的区别相同。

Before we turn to the readings' discussion questions, use this table as a compact decision aid.

转向阅读讨论问题之前，请把这张表当作简明决策辅助。

Exact untouched pixels suggest a role for masks and compositing.

如果未编辑像素必须精确不变，可以考虑蒙版与合成。

A new visual language may benefit from reference conditioning.

新的视觉语言，可能受益于参考图条件控制。

A reusable learned concept may motivate adaptation.

可复用的学习概念，可能需要模型适配。

A transformation that must survive motion needs a video workflow and temporal evidence.

必须经受运动的变换，需要视频工作流程和时间证据。

These approaches can cooperate in the same artwork.

这些方法可以在同一件作品中协作。

Choose according to the contract, then inspect the failure most likely to undermine it.

根据约定选择，再检查最可能破坏约定的失败。

That leaves us with a larger question for the readings: once production becomes easier, how do we choose a worthwhile direction and retain meaningful judgment over the result?

这给阅读留下更大的问题：制作更容易之后，如何选择值得追求的方向，并对结果保留有意义的判断？

An Alien Mind distinguishes achieving an assigned goal from generalizing human values in unfamiliar circumstances.

《An Alien Mind》区分了完成指定目标，与在陌生情境中泛化人类价值观。

Pachocki also raises questions about monitoring increasingly capable systems.

Pachocki 也提出了如何监测日益强大系统的问题。

Treat the essay as an argument containing claims and forecasts that we can examine.

请把文章视为包含主张与预测的论证，我们可以对其检视。

For our discussion, imagine an exhibition team delegating production while retaining responsibility for what the work communicates.

讨论时，想象一个展览团队委托制作，但仍对作品传达的内容负责。

Discuss the three questions on screen: explain the goal-and-values distinction, identify a forecast and the evidence it would need, and defend a boundary for human control.

讨论屏幕上的三个问题：解释目标与价值观的区别，找出一个预测及所需证据，并为人类控制的边界辩护。

Pause the video here.

请在这里暂停视频。

Our art-direction example is an analogy; it does not give an image editor the same agency or risk profile as an autonomous research system.

艺术指导只是类比；它并不赋予图像编辑器与自主研究系统相同的能动性或风险特征。

Koe's title makes a dramatic promise: How to fix your entire life in one day.

Koe 的标题作出了强烈承诺：《如何在一天内修复你的整个生活》。

The essay proposes examining identity and goals, interrupting habitual behavior, and turning reflection into action.

文章建议审视身份与目标、打断习惯行为，并把反思转化为行动。

We can discuss the usefulness of that proposal without accepting the title as a guarantee.

我们可以讨论这些建议是否有用，而不必把标题当作保证。

Imagine an artist who can generate a hundred attractive images but cannot choose what to make.

想象一个艺术家，能生成一百张好看的图像，却无法决定要创作什么。

Does easy production help clarify a direction, or make avoiding the decision easier?

轻松制作会帮助明确方向，还是让逃避决定更容易？

Discuss the three questions on screen.

请讨论屏幕上的三个问题。

Choose an idea worth using or challenging, explain the reason, and connect it to creative intention.

选一个值得采用或质疑的观点，解释理由，并把它与创作意图联系起来。

Pause the video here.

请在这里暂停视频。

You can use a fictional artist or public example; no personal disclosure is needed.

可以使用虚构艺术家或公开案例，不需要披露个人经历。

Runway's reading proposes approving an edited image before applying its appearance through a video.

Runway 的阅读提出：先批准编辑图像，再把它的外观应用到整段视频。

It offers a concrete answer to a communication problem: an art director can point to the desired look, rather than describe every feature in words.

它具体回答了一个沟通问题：艺术总监可以指出目标外观，而不必用文字描述每个特征。

Now bring back the column.

现在，让柱子重新出现。

A convincing keyframe does not tell us what the pear will look like after it reappears.

可信关键帧并不能告诉我们，梨再次显露后会是什么样子。

Discuss the three questions on screen: what the frame establishes, which claim deserves a harder test, and how the workflow serves AFTER RAIN's intention.

讨论屏幕上的三个问题：这一帧确定了什么，哪个主张需要更严格测试，以及流程如何服务于《雨后》的意图。

Pause the video.

请暂停视频。

The source is a vendor announcement; our job is to distinguish a useful demonstrated workflow from a reliability claim still needing evaluation.

来源是厂商公告；我们的任务是区分有用的已演示流程，与仍需评价的可靠性主张。

Return to your first judgment of the glass pear.

回到你最初对玻璃梨的判断。

We began with a small request and discovered that it touched the scene's physics, the artwork's intention, and the audience's experience over time.

我们从一个小要求出发，发现它触及场景物理、作品意图，以及观众随时间展开的体验。

You now have more precise ways to say what should change, what should survive, and how to judge the result.

现在，你有更精确的方法说明什么应改变、什么应保留，以及如何判断结果。

Finish with the three questions on screen.

最后，请讨论屏幕上的三个问题。

Where would you place the boundary between control and surprise?

你会把控制与惊喜之间的边界放在哪里？

How would you balance technical success and artistic purpose?

你会如何平衡技术成功与艺术目的？

What would you delegate, and what evidence would you require?

你会委托什么，又要求什么证据？

Pause the video for the final discussion.

请暂停视频，进行最终讨论。

Connect one mechanism or reading to a concrete artistic decision.

把一种机制或一篇阅读，与具体艺术决定联系起来。

Listen for an answer that makes you revise your own.

留意一个能让你修正自己看法的回答。

That revision is a fitting last act for a class about editing.

对一堂关于编辑的课来说，这样的修正，是恰当的最后一笔。