1
00:00:00,300 --> 00:00:04,220
Suppose the prompt asks for a blue cup with the word hello on a
sign.
假设提示要求一个蓝色杯子，牌子上写着 HELLO。

2
00:00:04,920 --> 00:00:10,120
Our original vector drawing shows a plausible cup, but the sign
reads H E L O.
我们的原创矢量图画出了看起来合理的杯子，但牌子写的是 HELO。

3
00:00:11,060 --> 00:00:16,580
A single preference score can tell us this output is worse, yet
it does not tell an editor exactly what to repair.
单个偏好分数可以表示输出较差，却不能准确告诉编辑器应该修复什么。

4
00:00:17,540 --> 00:00:22,320
Point to the inscription. Name the defect: a prompt-conditioned
lettering mismatch.
指向文字区域，说明缺陷类型：与提示不符的文字。

5
00:00:22,920 --> 00:00:25,020
Explain it: one letter L is missing.
解释原因：少了一个字母 L。

6
00:00:25,920 --> 00:00:27,640
Then state its priority for this task.
然后说明它在这个任务中的优先级。

7
00:00:28,500 --> 00:00:33,880
The example illustrates a diagnosis; it is not an output from a
trained detector or an annotated benchmark image.
这个示例是在展示诊断，并不是训练后检测器的输出，也不是基准中的标注图像。

8
00:00:34,580 --> 00:00:37,920
Structured Defect Grounding represents each defect as a tuple:
结构化缺陷定位把每个缺陷表示为一个元组：

9
00:00:38,320 --> 00:00:43,220
a location box, a defect type, a natural-language reason, and an
importance score.
位置框、缺陷类型、自然语言原因，以及重要性分数。

10
00:00:43,900 --> 00:00:46,480
An image can contain a variable number of these instances,
一张图像可以包含不同数量的缺陷实例，

11
00:00:46,780 --> 00:00:50,279
so the model predicts a set rather than one fixed image-level
label.
因此模型预测集合，而不是固定图像级标签。

12
00:00:51,000 --> 00:00:52,900
The fields stay attached to the same instance.
这些字段始终属于同一个缺陷实例。

13
00:00:53,540 --> 00:00:55,940
That matters when an image has several unrelated problems.
当图像存在多个无关问题时，这一点很重要。

14
00:00:56,660 --> 00:01:01,160
A malformed hand and a missing requested object need their own
locations and explanations,
变形的手与缺少提示物体，需要各自位置和解释，

15
00:01:01,480 --> 00:01:05,140
instead of a heatmap whose bright areas are disconnected from
the reasons for judging them.
而非只有亮区、却没将原因绑定到对应区域的热力图。

16
00:01:05,613 --> 00:01:08,393
The paper distinguishes artifacts from misalignments.
论文区分视觉瑕疵与提示不符。

17
00:01:08,873 --> 00:01:13,573
An artifact is an image-intrinsic visual flaw, such as
implausible geometry.
视觉瑕疵是图像本身的视觉问题，例如不合理的几何形状。

18
00:01:14,213 --> 00:01:17,273
A misalignment is a semantic error relative to the prompt.
提示不符则是相对提示的语义错误。

19
00:01:17,853 --> 00:01:23,293
Neatly written H E L O can therefore still be wrong when the
requested word is hello.
因此，即使 HELO 写得很整齐，在要求 HELLO 时它仍然不对。

20
00:01:24,393 --> 00:01:26,633
Importance is also attached to each defect.
重要性也属于每个缺陷。

21
00:01:27,233 --> 00:01:31,393
Its rubric considers visual prominence, semantic impact,
affected area,
评分规则考虑视觉显著性、语义影响与受影响面积，

22
00:01:31,873 --> 00:01:34,193
and whether the problem lies on the focal subject.
以及问题是否位于主体焦点。

23
00:01:35,033 --> 00:01:37,992
The dataset uses integer scores from one to one hundred.
数据集使用一到一百的整数分数。

24
00:01:38,553 --> 00:01:42,633
Our toy only names a priority; it does not invent a measured
importance number.
我们的示例只说明优先级，不虚构测量得到的重要性数值。

25
00:01:43,413 --> 00:01:50,193
To study this representation, the dataset contains thirty
thousand ninety-six images from four text-to-image generators.
为研究这种表示，数据集包含来自四个文生图生成器的三万零九十六张图像。

26
00:01:50,973 --> 00:01:55,473
Human annotators draw defect boxes, assign types, and write
short descriptions.
人工标注者绘制缺陷框、指定类型，并撰写短描述。

27
00:01:56,313 --> 00:02:01,153
A model then expands descriptions, supplies reasoning traces,
and scores importance.
模型随后扩展描述、补充推理轨迹，并评分重要性。

28
00:02:02,013 --> 00:02:05,433
The supervision is a combination of human and model annotations.
因此，监督信息由人工与模型标注共同组成。

29
00:02:06,333 --> 00:02:11,473
The split is prompt-disjoint: twenty-eight thousand nine hundred
forty-five training images
划分按提示互斥：二万八千九百四十五张训练图像，

30
00:02:11,533 --> 00:02:14,133
and one thousand one hundred fifty-one test images.
以及一千一百五十一张测试图像。

31
00:02:14,833 --> 00:02:16,313
No prompt appears in both splits.
没有提示同时出现在两个划分中。

32
00:02:17,153 --> 00:02:20,713
That gives the reported detector evaluation a defined held-out
setting,
这为检测评估提供明确的留出设置，

33
00:02:21,213 --> 00:02:24,033
rather than testing on the same prompt records used for
training.
而不是用训练时同样的提示记录测试。

34
00:02:24,660 --> 00:02:28,780
The detector starts from Qwen three V L with four billion
parameters.
检测器从四十亿参数的 Qwen3-VL 开始。

35
00:02:29,580 --> 00:02:33,840
Supervised fine-tuning teaches the required reasoning and
structured defect output.
监督微调教授所需的推理与结构化缺陷输出。

36
00:02:34,660 --> 00:02:41,260
A second policy-optimization stage rewards spatial accuracy,
description consistency, and importance estimation.
第二个策略优化阶段奖励空间准确性、描述一致性和重要性估计。

37
00:02:42,360 --> 00:02:44,500
Those rewards sit behind a format gate.
这些奖励之前设置了格式门控。

38
00:02:45,080 --> 00:02:50,000
The answer must contain a parseable defect list and
geometrically valid boxes
答案必须包含可解析缺陷列表与几何有效的框，

39
00:02:50,060 --> 00:02:52,300
before it receives the normal composite reward.
才可获得正常的组合奖励。

40
00:02:53,180 --> 00:02:54,980
A valid answer still needs to be correct:
格式有效的答案仍需要正确：

41
00:02:55,600 --> 00:03:01,060
locating a box is different from explaining the defect, and both
are different from estimating its impact.
定位框、解释缺陷与估计影响，是不同的问题。

42
00:03:01,627 --> 00:03:05,407
Evaluation separates image-level and defect-level questions.
评估区分图像级与缺陷级问题。

43
00:03:06,167 --> 00:03:11,427
Image-level metrics check defect-type presence, including images
with no reference instance of that type.
图像级指标检查缺陷类型是否存在，也包括未标注该类型实例的图像。

44
00:03:12,287 --> 00:03:16,607
Defect-level metrics match predictions to annotated instances of
the same type,
缺陷级指标将预测与同类型标注实例匹配，

45
00:03:17,067 --> 00:03:20,047
then evaluate boxes, descriptions, and importance.
再评价框、描述和重要性。

46
00:03:21,027 --> 00:03:26,247
For the next chart, the question is artifact localization on
images with annotated artifacts.
下一张图考察存在标注视觉瑕疵的图像上的定位能力。

47
00:03:27,247 --> 00:03:33,687
Box F one at zero point five uses an overlap threshold of zero
point five when evaluating predicted boxes.
BoxF1@0.5 在评价预测框时使用零点五的重叠阈值。

48
00:03:34,327 --> 00:03:38,547
It is a localization F one score, not an overall
image-preference rating
它是定位 F1 分数，不是整体图像偏好评分，

49
00:03:38,647 --> 00:03:40,767
or a promise that half the defects will be repaired.
也不是承诺一半缺陷能被修复。

50
00:03:41,400 --> 00:03:46,320
On the test set, zero-shot Gemini three Pro scores zero point
two zero zero
测试集上，零样本 Gemini 3 Pro 得分为零点二零零，

51
00:03:46,380 --> 00:03:48,840
on artifact Box F one at zero point five.
该指标是视觉瑕疵 BoxF1@0.5。

52
00:03:49,420 --> 00:03:52,880
The supervised SDG detector scores zero point two five five;
监督训练 SDG 检测器为零点二五五；

53
00:03:53,220 --> 00:03:56,560
its policy-optimized variant scores zero point two six three.
策略优化变体为零点二六三。

54
00:03:57,319 --> 00:04:03,380
This is a task-trained four-billion-parameter detector compared
with a zero-shot general vision-language baseline,
这是针对任务训练的四十亿参数检测器与零样本通用视觉语言基线的比较，

55
00:04:03,440 --> 00:04:04,680
on this localization metric.
比较针对这一定位指标。

56
00:04:05,560 --> 00:04:09,300
Independent human reannotation is zero point two seven eight in
the same table.
同一表格中，独立人工重新标注的参考值为零点二七八。

57
00:04:10,080 --> 00:04:15,840
The comparison is useful evidence about grounding; it is not a
ranking of every model on every image-quality task.
该比较为定位能力提供证据，并不是所有模型在所有图像质量任务上的排名。

58
00:04:16,600 --> 00:04:18,900
One use is generator post-training.
一种用途是生成器后训练。

59
00:04:19,620 --> 00:04:25,320
BoxFlow G R P O turns detected regions and their importance into
spatial reward penalties,
BoxFlow-GRPO 把区域及其重要性转为空间奖励惩罚，

60
00:04:25,840 --> 00:04:29,640
then normalizes advantages at each location across a prompt
group.
再在提示组内部逐位置归一化优势。

61
00:04:30,340 --> 00:04:34,840
A high-impact problem can receive stronger local pressure than
an unrelated part of the image.
影响较大的问题区域可以获得比无关图像部分更强的局部训练压力。

62
00:04:35,780 --> 00:04:36,920
Another use is editing.
另一种用途是编辑。

63
00:04:37,620 --> 00:04:42,700
With the same image editor, structured boxes and text are
compared against caption-only feedback
使用相同编辑器，将结构化框与文本和仅描述反馈比较，

64
00:04:42,760 --> 00:04:44,880
on eight hundred seventy-three valid samples.
比较使用八百七十三个有效样本。

65
00:04:45,500 --> 00:04:48,959
Against that baseline, ten point three one percent favor SDG,
相对该基线，百分之十点三一偏好 SDG，

66
00:04:49,560 --> 00:04:51,640
two point seven five percent favor the baseline,
百分之二点七五偏好基线，

67
00:04:52,040 --> 00:04:54,540
and eighty-six point nine four percent are ties.
百分之八十六点九四是平局。

68
00:04:55,120 --> 00:04:56,060
Return to our sign.
回到牌子。

69
00:04:56,680 --> 00:04:58,620
A box alone says where to look.
框告诉我们看哪里。

70
00:04:59,140 --> 00:05:01,800
A reason says why the lettering violates the prompt.
原因说明文字为何违背提示。

71
00:05:02,540 --> 00:05:05,780
Type distinguishes a prompt mismatch from a visual artifact.
类型区分提示不符与视觉瑕疵。

72
00:05:06,480 --> 00:05:09,720
Importance tells the next system which issue deserves attention.
重要性告诉后续系统哪些问题值得关注。

73
00:05:10,500 --> 00:05:12,740
The information is useful because it stays together.
这些信息之所以有用，是因为它们始终绑定在一起。

74
00:05:13,360 --> 00:05:15,960
The central idea is a richer interface for feedback:
核心思想是更丰富的反馈接口：

75
00:05:16,520 --> 00:05:21,460
not only whether an image is good, but which instance is wrong
and what needs to change.
不仅问图像好不好，还指出哪个实例错误，以及需要改变什么。

76
00:05:22,340 --> 00:05:24,900
Structured diagnosis does not itself edit the cup.
结构化诊断本身不会编辑杯子。

77
00:05:25,560 --> 00:05:29,580
It gives an editor or a training objective a more specific
problem to act on.
它为编辑器或训练目标提供更明确的行动问题。