Flow-based image generators achieve strong semantic synthesis but still produce localized structural distortions, such as malformed hands and limbs. Existing post-training methods typically optimize global scalar rewards, which provide weak spatial credit assignment and may alter well-formed regions while correcting local defects. We present Diagnose-to-Generate (D2G), a grounded alignment framework that separates visual diagnosis from generative optimization. We construct Achilles-600K, a region-annotated dataset for structural anomalies, and train ARGUS through supervised cold-start and Region-Aware Group Relative Policy Optimization to predict anomaly categories and bounding boxes. These diagnoses are converted into confidence-weighted spatial rewards for our Patch-level Flow-GRPO. We further identify that global stochastic exploration can shift semantic regions across group rollouts, making patch-wise relative advantages spatially inconsistent. Our Trajectory-Preserving Stochastic Sampling (TPSS) anchors early denoising with deterministic ODE trajectories and introduces stochasticity only at later steps, thereby limiting intra-group spatial drift while retaining local exploration. Experiments on FLUX.1-dev and SD3.5 show improved structural fidelity with preserved FID and CLIP scores. Consistent gains under HumanRefiner, zero-shot GPT-4.1, and blinded human evaluation further demonstrate that the improvements extend beyond the training reward model.
Figures & tables
Figure 1. Illustration of global and grounded structural alignment. (a) An image-level reward does not identify the region responsible for a structural defect. (b) D2G uses ARGUS diagnoses to construct spatial rewards, concentrating structural feedback around anomalous regions while retaining a global quality objective.
Figure 2. Achilles-600k construction and benchmark statistics. Diverse generators provide images for human-verified structural annotation, from which 10K grounded Reverse-CoT instances are selected for diagnostic cold-start training. The Achilles-Benchmark is independently curated and covers diverse anomaly categories, object scales, and numbers of localized defects.
Figure 3. Overview of Diagnose-to-Generate (D2G). Stage I trains ARGUS via grounded SFT and Region-Aware GRPO to localize structural anomalies. Stage II uses ARGUS diagnoses to spatially supervise Patch-level Flow-GRPO. TPSS preserves rollout correspondence, enabling localized advantages to update diagnosed regions while limiting changes elsewhere.
Figure 4. Qualitative comparison of generative alignment. (a) Global Flow-GRPO alters surrounding appearance, whereas D2G preserves it during local correction. (b) Update magnitudes are globally distributed for Flow-GRPO but concentrated around ARGUS-diagnosed regions for D2G.
Category
Model
L2 Labels Macro
Artifact Detection
Spatial
Human Align.
Prec.
Rec.
F1
Prec.
Rec.
F1
mIoU ( ↑ )
Acc. (%) ( ↑ )
Proprietary APIs
GPT-4.1 † ( OpenAI, 2025a )
0.4452
0.2105
0.2855
0.6315
0.4055
0.4950
0.2361
68.5
Gemini 3.1 Pro † ( Google, 2025 )
0.4294
0.3572
0.3905
0.5577
0.8214
0.6631
0.2814
72.4
Zero-Shot Foundation
Qwen2.5-VL-7B ( Bai et al., 2025b )
0.3395
0.0995
0.1538
0.5674
0.3597
0.4403
–
55.2
InternVL3.5-8B ( Wang et al., 2025b )
0.3802
0.1205
0.1830
0.5810
0.3822
0.4611
–
58.1
Qwen3-VL-8B ( Bai et al., 2025a )
0.4105
0.1620
0.2325
0.6015
0.4105
0.4880
–
61.3
Table 1. Diagnostic performance on Achilles-Benchmark. L2 Macro averages precision, recall, and F1 over four anomaly categories; Binary evaluates anomaly detection, and mIoU evaluates localization. † denotes structured CoT prompting for box outputs. Human Align. follows ( Wang et al., 2026 ) .
Category
Model
Abnormal Interaction
Human Anatomy
Animal Anatomy
Object Morphology
Prec.
Rec.
F1
Prec.
Rec.
F1
Prec.
Rec.
F1
Prec.
Rec.
F1
Proprietary APIs
GPT-4.1 ( OpenAI, 2025a )
0.3112
0.2054
0.2478
0.7305
0.3415
0.4655
0.5840
0.2512
0.3512
0.2645
0.1130
0.1583
Gemini 3.1 Pro ( Google, 2025 )
0.3204
0.3015
0.3106
0.7645
0.5012
0.6054
0.5620
0.3840
0.4562
0.2215
0.4420
0.2885
Zero-Shot Found.
Qwen2.5-VL-7B ( Bai et al., 2025b )
0.2008
0.1180
0.1486
0.6182
0.1358
0.2227
0.4361
0.1005
0.1634
0.1028
0.0436
0.0612
InternVL3.5-8B ( Wang et al., 2025b )
0.2315
0.1350
0.1705
0.6410
0.1620
0.2586
0.4720
0.1245
0.1965
0.1345
0.0620
0.0848
Qwen3-VL-8B ( Bai et al., 2025a )
0.2550
0.1640
0.1988
0.6855
0.2105
0.3222
0.5105
0.1660
0.2505
0.1802
0.0915
0.1215
Table 2. Diagnostic performance across fine-grained distortion categories. We report precision, recall, and F1 for abnormal interaction, human anatomy, animal anatomy, and object morphology. ARGUS consistently improves category-level F1 across three underlying VLM backbones.
Model / Alignment Strategy
Structural Fidelity on Achilles-Benchmark ( ↑ )
Global Preservation and Human Alignment
Interact.
Human
Animal
Object
Overall
FID-30K ( ↓ )
CLIP ( ↑ )
Human Win ( ↑ )
Panel A: Contemporary Foundation Models (Zero-Shot)
GPT-image-1 ( Chen et al., 2025 )
81.54
45.00
49.00
91.41
63.08
–
–
52.4%
HunyuanVideo ( Kong et al., 2024 )
86.15
43.20
48.50
88.50
62.15
–
–
51.0%
Bagel ( Deng et al., )
85.96
41.00
48.00
87.92
60.53
–
–
53.1%
Seedance 1.0 Pro ( Gao et al., 2025 )
78.37
41.00
43.00
90.56
59.54
–
–
48.5%
Table 3. End-to-end structural alignment and independent evaluation. Panels A–B report Achilles-Benchmark structural fidelity and global preservation. Panel C evaluates FLUX.1-dev outputs using HumanRefiner, zero-shot GPT-4.1, and double-blind human preference, and includes local alignment baselines. Panel D evaluates cross-architecture transfer on SD3.5-M. Best and second-best results are marked within each panel.
Ablation Variant
Diag. F1
mIoU
Fidelity
FID-30K
Stage I: Diagnostic Specialist
Full ARGUS
0.8321
0.7540
–
–
w/o Reverse-CoT
0.7510
0.6150
–
–
w/o Bipartite GIoU Match
0.6845
0.5210
–
–
w/o Consistency Gate ( rc )
0.7890
0.4250
–
–
w/o Multi-Bucket Sampling
0.7415
0.7105
–
–
Table 4. Component ablations. Stage I reports diagnostic F1 and localization mIoU; Stage II reports Achilles structural fidelity and FID-30K. Reverse-CoT evaluates the grounded data engine, while Confidence-Decay evaluates robustness to uncertain ARGUS diagnoses.
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
Figure S1 . Examples of diagnostic feedback. Image-level scores summarize overall quality, while textual assessments can describe defects without locating them. ARGUS additionally predicts anomaly boxes, making its feedback usable for spatial reward assignment. The examples illustrate output formats rather than a general impossibility of grounding with other MLLMs.
Setting
Value
Stage I: SFT
Training examples
10K
Global batch size
128
Training epochs
2
Peak learning rate
2×10−5
Learning-rate schedule
Cosine
Appendix
Table S1 . Training configuration. Gaussian smoothing is specified in latent-grid units. The Stage-I localization threshold and Stage-II confidence threshold serve different purposes.
Model
Prec.
Rec.
mIoU
GPT-4.1 ( OpenAI, 2025a )
0.6315
0.4055
0.2361
Gemini 3.1 Pro ( Google, 2025 )
0.5577
0.8214
0.2814
ARGUS
0.8655
0.8012
0.7540
Appendix
Table S2 . Forced-box diagnostic evaluation. Structured CoT prompting requires proprietary models to emit bounding boxes. Precision and recall refer to binary anomaly detection.
Alignment
Human Fidelity
Overall Fidelity
FID-30K
Base SD3.5-M
39.33
54.82
20.15
Global Flow-GRPO
42.10
58.20
26.40
Patch-level Flow-GRPO
76.45
81.50
20.10
Appendix
Table S3 . Stage-II alignment on SD3.5-Medium. Fidelity is evaluated on Achilles-Benchmark. Higher fidelity and lower FID are preferable.
Figure S2 . Sensitivity to the TPSS branching timestep. The original sweep reports structural fidelity and FID at different values of tcritical . Intermediate thresholds provide a better trade-off than nearly deterministic sampling or early stochastic exploration.
Variant
Diag. F1
mIoU
Fidelity
FID
Full D2G
0.8321
0.7540
85.60
18.35
w/o Reverse-CoT
0.751
0.615
–
–
w/o Confidence-Decay
–
–
84.15
21.40
Appendix
Table S4 . Additional pipeline ablations. Reverse-CoT is evaluated through Stage-I diagnostics; Confidence-Decay is evaluated through Stage-II generation. Precision is retained as reported in the corresponding experiments.