1
00:00:00,300 --> 00:00:04,220
Suppose the prompt asks for a blue cup with the word hello on a
sign.

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.

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.

4
00:00:17,540 --> 00:00:22,320
Point to the inscription. Name the defect: a prompt-conditioned
lettering mismatch.

5
00:00:22,920 --> 00:00:25,020
Explain it: one letter L is missing.

6
00:00:25,920 --> 00:00:27,640
Then state its priority for this task.

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.

8
00:00:34,580 --> 00:00:37,920
Structured Defect Grounding represents each defect as a tuple:

9
00:00:38,320 --> 00:00:43,220
a location box, a defect type, a natural-language reason, and an
importance score.

10
00:00:43,900 --> 00:00:46,480
An image can contain a variable number of these instances,

11
00:00:46,780 --> 00:00:50,279
so the model predicts a set rather than one fixed image-level
label.

12
00:00:51,000 --> 00:00:52,900
The fields stay attached to the same instance.

13
00:00:53,540 --> 00:00:55,940
That matters when an image has several unrelated problems.

14
00:00:56,660 --> 00:01:01,160
A malformed hand and a missing requested object need their own
locations and explanations,

15
00:01:01,480 --> 00:01:05,140
instead of a heatmap whose bright areas are disconnected from
the reasons for judging them.

16
00:01:05,613 --> 00:01:08,393
The paper distinguishes artifacts from misalignments.

17
00:01:08,873 --> 00:01:13,573
An artifact is an image-intrinsic visual flaw, such as
implausible geometry.

18
00:01:14,213 --> 00:01:17,273
A misalignment is a semantic error relative to the prompt.

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.

20
00:01:24,393 --> 00:01:26,633
Importance is also attached to each defect.

21
00:01:27,233 --> 00:01:31,393
Its rubric considers visual prominence, semantic impact,
affected area,

22
00:01:31,873 --> 00:01:34,193
and whether the problem lies on the focal subject.

23
00:01:35,033 --> 00:01:37,992
The dataset uses integer scores from one to one hundred.

24
00:01:38,553 --> 00:01:42,633
Our toy only names a priority; it does not invent a measured
importance number.

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.

26
00:01:50,973 --> 00:01:55,473
Human annotators draw defect boxes, assign types, and write
short descriptions.

27
00:01:56,313 --> 00:02:01,153
A model then expands descriptions, supplies reasoning traces,
and scores importance.

28
00:02:02,013 --> 00:02:05,433
The supervision is a combination of human and model annotations.

29
00:02:06,333 --> 00:02:11,473
The split is prompt-disjoint: twenty-eight thousand nine hundred
forty-five training images

30
00:02:11,533 --> 00:02:14,133
and one thousand one hundred fifty-one test images.

31
00:02:14,833 --> 00:02:16,313
No prompt appears in both splits.

32
00:02:17,153 --> 00:02:20,713
That gives the reported detector evaluation a defined held-out
setting,

33
00:02:21,213 --> 00:02:24,033
rather than testing on the same prompt records used for
training.

34
00:02:24,660 --> 00:02:28,780
The detector starts from Qwen three V L with four billion
parameters.

35
00:02:29,580 --> 00:02:33,840
Supervised fine-tuning teaches the required reasoning and
structured defect output.

36
00:02:34,660 --> 00:02:41,260
A second policy-optimization stage rewards spatial accuracy,
description consistency, and importance estimation.

37
00:02:42,360 --> 00:02:44,500
Those rewards sit behind a format gate.

38
00:02:45,080 --> 00:02:50,000
The answer must contain a parseable defect list and
geometrically valid boxes

39
00:02:50,060 --> 00:02:52,300
before it receives the normal composite reward.

40
00:02:53,180 --> 00:02:54,980
A valid answer still needs to be correct:

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.

42
00:03:01,627 --> 00:03:05,407
Evaluation separates image-level and defect-level questions.

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.

44
00:03:12,287 --> 00:03:16,607
Defect-level metrics match predictions to annotated instances of
the same type,

45
00:03:17,067 --> 00:03:20,047
then evaluate boxes, descriptions, and importance.

46
00:03:21,027 --> 00:03:26,247
For the next chart, the question is artifact localization on
images with annotated artifacts.

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.

48
00:03:34,327 --> 00:03:38,547
It is a localization F one score, not an overall
image-preference rating

49
00:03:38,647 --> 00:03:40,767
or a promise that half the defects will be repaired.

50
00:03:41,400 --> 00:03:46,320
On the test set, zero-shot Gemini three Pro scores zero point
two zero zero

51
00:03:46,380 --> 00:03:48,840
on artifact Box F one at zero point five.

52
00:03:49,420 --> 00:03:52,880
The supervised SDG detector scores zero point two five five;

53
00:03:53,220 --> 00:03:56,560
its policy-optimized variant scores zero point two six three.

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,

55
00:04:03,440 --> 00:04:04,680
on this localization metric.

56
00:04:05,560 --> 00:04:09,300
Independent human reannotation is zero point two seven eight in
the same table.

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.

58
00:04:16,600 --> 00:04:18,900
One use is generator post-training.

59
00:04:19,620 --> 00:04:25,320
BoxFlow G R P O turns detected regions and their importance into
spatial reward penalties,

60
00:04:25,840 --> 00:04:29,640
then normalizes advantages at each location across a prompt
group.

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.

62
00:04:35,780 --> 00:04:36,920
Another use is editing.

63
00:04:37,620 --> 00:04:42,700
With the same image editor, structured boxes and text are
compared against caption-only feedback

64
00:04:42,760 --> 00:04:44,880
on eight hundred seventy-three valid samples.

65
00:04:45,500 --> 00:04:48,959
Against that baseline, ten point three one percent favor SDG,

66
00:04:49,560 --> 00:04:51,640
two point seven five percent favor the baseline,

67
00:04:52,040 --> 00:04:54,540
and eighty-six point nine four percent are ties.

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

69
00:04:56,680 --> 00:04:58,620
A box alone says where to look.

70
00:04:59,140 --> 00:05:01,800
A reason says why the lettering violates the prompt.

71
00:05:02,540 --> 00:05:05,780
Type distinguishes a prompt mismatch from a visual artifact.

72
00:05:06,480 --> 00:05:09,720
Importance tells the next system which issue deserves attention.

73
00:05:10,500 --> 00:05:12,740
The information is useful because it stays together.

74
00:05:13,360 --> 00:05:15,960
The central idea is a richer interface for feedback:

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.

76
00:05:22,340 --> 00:05:24,900
Structured diagnosis does not itself edit the cup.

77
00:05:25,560 --> 00:05:29,580
It gives an editor or a training objective a more specific
problem to act on.