Standard clothing asset generation---restoring forward-facing flat-lay garment images from diverse real-world contexts---holds immense commercial value yet demands both macroscopic topological accuracy and microscopic physical fidelity. Although our previous work RAGDiffusion effectively eradicated large-scale structural hallucinations via retrieval-augmented macro-constraints, achieving industrial-grade micro-texture realism remains an unsolved bottleneck. We formally identify this limitation as High-Frequency Trajectory Collapse: supervised fine-tuning (SFT) converges to the conditional mean of the training distribution, which is dominated by smooth, low-frequency textures, causing high-frequency patterns (e.g., fabric weaves, intricate logos) to become nearly un-sampleable. Naively applying Reinforcement Learning (RL) post-training further triggers Artifact Hacking, where models exploit semantic biases in generic reward models by generating deceptive checkerboard noise. Our key insight is that RL can fundamentally reshape the sampling distribution of flow models---elevating the probability of high-fidelity trajectories under accurate reward guidance---while adversarial regularization prevents exploitation of reward blind spots. Realizing this principle requires three prerequisites: (i)inherent capacity, established through a 27,725-pair high-complexity garment dataset (STGarment-Plus) and a Dual-Image-Stream FLUX architecture upgrade; (ii)perceptive reward, provided by a novel attribute-aware reward model (Garment-RM) trained on 500K images via fine-grained contrastive learning, achieving 84.67% human preference accuracy; and (iii)hacking prevention, enforced by our Adversarial-Regularized GRPO (AR-GRPO) strategy that integrates a dynamic discriminator into the RL sampling trajectory to penalize artifacts while enriching authentic high-frequency details.
Figures & tables
Fig. 1: High-Frequency Trajectory Enhancement via Distribution Reshaping. Left: Current SFT-based models regress to the conditional mean, smoothing out textures. Right: RL elevates probability in true high-frequency modes, enabling various and high-fidelity sampling.
Fig. 2: Breaking the SFT saturation bottleneck with AR-GRPO. During the SFT phase, the generative model reaches a performance plateau, while extending the training duration or scaling the data yields diminishing returns. The introduction of AR-GRPO triggers a sharp decline in FID, successfully breaking this bottleneck. The visual comparisons demonstrate SFT’s limitations in capturing high-frequency physical details (e.g., complex fabric textures and intricate patterns) and the significant enhancements brought by RL optimization.
Fig. 3: (a) shows Frequency-domain comparison between the base (SFT) model generation DBase , AR-GRPO results DRL , and real images DReal . Compared with the Base model, RL yields lower spectral discrepancy and weaker directional bias. (b) shows radially averaged spectral error curves with respect to the real-image spectrum. The RL model shows lower error over a broad middle-to-high frequency range.
Fig. 4: Overall framework of our previous work RAGDiffusion. StructureNet provides latent structural embeddings, while SLLE facilitates the embedding fusion along with landmark retrieval. The generative model synthesizes multiple conditions to achieve omni-level high-fidelity generation.
Fig. 5: AR-GRPO training framework consists of four phases: Phase 1 generates diverse trajectories via SDE sampling; Phase 2 trains the discriminator hϕ adversarially to establish an authenticity reward; Phase 3 computes combined advantages from Garment-RM, LPIPS, and discriminator; Phase 4 updates the generative policy. This adversarial regularization ensures the model achieves high fidelity while avoiding artifacts.
Fig. 6: Overview of STGarment-Plus. (a) and (b) give a visual comparison of STGarment-Plus and open-source datasets. (c) STGarment-Plus covers challenging real-world scenes, diverse garment styles, and difficult high-frequency textures.
Fig. 7: Dual-Image-Stream DiT framework. Our model introduces a dedicated conditioning image stream by replicating part of the denoising image stream and its pretrained weights.
Fig. 8: (a) Adv-GRPO [ 58 ] aggressively updates all reward models with adversarial loss, causing them to lose classification ability. (b) Reward trajectories: the discriminator first accelerates early reward rise, then stabilizes both RGRM and RLPIPS at an artifact-free equilibrium. (c) Standard GRPO and Adv-GRPO collapse into checkerboard artifacts, while AR-GRPO maintains clean outputs. Best viewed when zoomed in.
Fig. 9: Qualitative comparison with state-of-the-art garment generation baselines. Best viewed when zoomed in.
Method
SSIM ↑
LPIPS ↓
FID ↓
KID ↓
DISTS ↓
SDXL Backbone (Trained on STGarment)
PBE
0.5278
0.4821
26.55
10.08
0.352
ControlNet
0.5225
0.4792
23.02
9.231
0.345
IP-Adapter
0.6067
0.4728
14.50
3.747
0.264
TryOffDiff
0.6124
0.5007
34.34
14.19
0.335
TryOffAnyone
0.6575
0.4335
18.86
7.919
0.245
TABLE I: Quantitative comparisons of generative performance against state-of-the-art baselines. The upper section evaluates SDXL-based models, while the lower section compares RAGDiffusion++ against FLUX-based and large-scale commercial models.
Method
SSIM ↑
LPIPS ↓
FID ↓
KID ↓
DISTS ↓
Base Model Components in RAGDiffusion
Vanilla Model
0.6576
0.3961
10.65
1.405
0.222
+ emb
0.6614
0.3917
10.55
1.354
0.217
+ emb + sil
0.6926
0.3690
10.22
0.940
0.210
RAGDiffusion
0.6963
0.3684
9.990
1.092
0.192
Data and Architecture Upgrades (SFT Phase)
TABLE II: Quantitative ablation study of network components, architecture and data upgrades, and AR-GRPO rewards.
Dataset
Viton-HD
DC-Upper
DC-Dress
Method
FID ↓
KID ↓
FID ↓
KID ↓
FID ↓
KID ↓
SDXL-based Models
PBE
80.8
52.9
45.0
21.1
105.1
74.6
ControlNet
69.9
43.2
48.9
23.6
118.3
95.4
IP-Adapter
46.8
17.8
28.8
11.1
47.8
19.0
TryOffDiff*
14.0
3.31
38.9
16.7
74.3
48.6
TABLE III: Cross-dataset evaluation on Viton-HD and DressCode (DC). Models marked with * denote inner-dataset testing (trained on the target dataset), while others are evaluated in a cross-dataset setting.
Figure 13
Fig. 12: Ablation of core reward components in AR-GRPO. By comparing (a) the absence of Garment-RM (resulting in blurred textures and local misalignment), (b) the absence of the discriminator Rdis (degrading to standard GRPO [ 34 ] and causing dense checkerboard artifacts), and (c) the use of a generic DINO [CLS] reward (leading to severe trajectory distortion), we prove the indispensability of specialized reward models and adversarial regularization in preserving fabric realism and preventing reward hacking. Best viewed when zoomed in.
Fig. 13: We observe that interpolation between different reward models can always get higher Top-1 accuracy. While Garment-RM achieves the best accuracy in a single reward setting, interpolating it with LPIPS score produces a bit better performance.
Fig. 14: T-SNE figures of Garment-RM embeddings and human evaluation on preference selection to validate the effectiveness of Garment-RM.
Model
Preference
Recall
Filter
Acc.
@1
@2
@4
@1
@2
@4
HPS v2.1
33.48
2.49
7.62
24.68
10.82
20.42
37.94
ImageReward
38.95
4.27
10.97
29.91
14.27
23.36
42.46
PickScore
45.04
6.91
15.79
39.00
19.86
31.59
49.16
MSE Score
52.25
20.47
36.92
72.12
9.40
16.40
36.06
CLIP Similarity
77.97
22.45
40.58
72.07
67.60
82.02
92.59
TABLE IV: Results of Garment-RM and comparison methods on human preference prediction. Preference accuracy and Recall/Filter scores are evaluated on a test set of 1,969 prompts (with 9 pairs per sample). All scores are averaged per prompt.
Fig. 15: Performance comparison of fine-grained attribute perception. Garment-RM comprehensively outperforms state-of-the-art large language models (GPT-4o and Qwen3-VL), leading to reliable, attribute-aware signals for the RL optimization phase.