Numerous distillation methods have been developed to accelerate diffusion models. Recent research indicates that hybrid strategies---blending different paradigms---yield superior results compared to single-method approaches. However, current works predominantly rely on a simple combination of objectives, leaving the question of how to effectively synergize them largely underexplored. In this paper, we specifically investigate the synergy between Score Distillation and Adversarial Distillation. We reveal that their operating zones are distinct and complementary: Score Distillation is effective in high-noise regimes for capturing the global distribution yet lacks fine-grained supervision in low-noise regions. Conversely, Adversarial Distillation is effective in low-noise regions but is prone to instability and suffers from discriminator failure when applied to high-noise intervals. Leveraging this insight, we propose Efficient Adversarial Score Distillation (EASD), a framework that synergizes these objectives via a regime-adaptive sampling strategy, directing each objective to focus on its effective interval. We further apply this paradigm to DiT-based Flow Matching models. Empirically, our method achieves a promising one-step FID of 1.39 on ImageNet 256×256 using the SiT-XL/2+REPA model. This work provides a concrete guideline for maximizing the potential of hybrid distillation. Our code and model are available at https://github.com/pf0607/EASD .
Figures & tables
Figure 1: Qualitative Results on ImageNet 256 × 256. Samples generated by our EASD-XL/2 with one inference step . More results are provided in Appendix Appendix I .
Figure 2
Figure 3: Visual Analysis of Adversarial Distillation Regimes. (Left) Low Noise ( t→0 ): Due to disjoint manifolds, the discriminator overfits, forcing the generator to collapse to few modes (red clump) ( Mode Collapse ). (Center) Effective Regime ( t≈Mid ): Noise injection functions as Manifold Smoothing , creating optimal overlap for gradient flow while preserving fine-grained details and roubust distribution . (Right) High Noise ( t→1 ): The distributions merge into an indistinguishable Gaussian ball, causing the discriminator to lose all guidance ( Signal Loss ).
Figure 4Table 5Table 6
Method
Adversarial Updates
Time
FID@100K ↓
FID@200K ↓
DMD2 [ 68 ]
Gen. + Disc. + Fake Score ∗
28h
1.72
1.52
+ Regime-Adaptive Sampling
Gen. + Disc. + Fake Score ∗
28h
1.50
1.35
EASD (Full)
Gen. + Disc. †
23h
1.46
1.26
Table 7: Strict Ablation with DMD2 on ImageNet- 64×64 (5:1 update ratio, batch size 280; all runs update Gen. + Fake Score in score distillation). ∗ Fake-score is jointly updated in discriminator optimization. † Frozen-teacher discriminator without updating fake-score.
Appendix figures & tables9 assets
Supplementary material from the paper’s appendix.
Appendix
Method (1-step SDXL)
COCO-10K FID ↓
CLIP Score ↑
SDXL-Turbo (ADD)
23.92
0.337
DMD2
19.01
0.336
EASD (Ours)
18.81
0.341
Appendix
Table 8: Comparison of 1-step generation on COCO-10K with SDXL-based methods.
Score Distillation
Adversarial Distillation
FID ↓
U(0.0,1.0)
U(0.0,1.0)
6.98
LN(0.8,1.6)
U(0.0,1.0)
5.58
U(0.0,1.0)
LN(−0.8,1.6)
6.06
LN(1.0,1.6)
LN(−1.0,1.6)
5.11
LN(0.8,1.6)
LN(−0.8,1.6)
5.02
LN(0.4,1.6)
LN(−0.4,1.6)
5.61
Appendix
Table 9: Ablation study of sampling strategy combinations for score distillation and adversarial distillation on CIFAR-10.
Method
Memory (BS=256)
Best-Practice Config & Memory
20-Epoch Time
FID (20ep) ↓
Best-Practice Time
Best FID ↓
PiFlow [ 5 ]
17.8 GB
BS=4096, 35.8 GB
6h 10min
5.13
138h
2.85
DMF [ 23 ]
56.4 GB
BS=256, 56.4 GB
11h 30min
2.79
46h
2.16
EASD (Ours)
48.7 GB
BS=256, 48.7 GB
8h 30min
1.47
26h
1.39
Appendix
Table 10: Efficiency comparison of distillation methods on SiT-XL/2+REPA [ 70 ] .
Figure 5: FID convergence curves.
Score Distillation
Adversarial Distillation
FID ↓
LN(0.8,1.6) , truncated to (0,0.2)
LN(−0.8,1.6) , truncated to (0.8,1.0)
1.47
LN(0.8,1.6) , truncated to (0,0.2)
LN(−0.8,1.6) , untruncated
1.50
LN(0.8,1.6) , untruncated
LN(−0.8,1.6) , truncated to (0.8,1.0)
1.49
Appendix
Table 11: Ablation study on truncated sampling strategies for score distillation and adversarial distillation.
Figure 6: Visual comparison of training dynamics with and without CFG. Rows correspond to samples generated by models trained without CFG (Top) and with CFG w=1.35 (Bottom) across different training steps (500 to 2000). The model utilizing CFG demonstrates rapid initial convergence, producing recognizable objects with sharp features. However, this acceleration leads to premature convergence, where the model collapses into a limited set of high-contrast modes. In contrast, the model without CFG evolves more gradually but preserves greater diversity, avoiding the oversaturated artifacts associated with recursive guidance.
Method
IS ↑
FID ↓
sFID ↓
Precision ↑
Recall ↑
w/o CFG
233.32
2.30
4.67
0.787
0.624
w/ CFG ( w=1.35 )
341.98
3.92
5.05
0.875
0.508
Appendix
Table 12: Quantitative Impact of CFG. Comparison of metrics for models trained with and without CFG. The use of CFG leads to higher Precision and Inception Score (IS) but significantly lower Recall and worse FID, indicating severe mode collapse.
Figure 7: Additional One-Step Samples on ImageNet 256×256 . Selected samples generated by EASD. The results demonstrate that our method achieves photorealistic quality and covers diverse modes across various categories.
Figure 8: Additional One-Step Samples on ImageNet 256×256 . Selected samples generated by EASD. The results demonstrate that our method achieves photorealistic quality and covers diverse modes across various categories.
Flow and diffusion models suffer from slow inference due to computationally expensive numerical integration. Distillation provides a promising way for a student model to learn from a teacher's dynamics, enabling one-step or few-step generation. However, existing methods often depend on curated distillation datasets, costly teacher rollouts, or auxiliary proxy networks, which complicate model training and scaling. In this work, we propose Consistent Distribution Matching, a simulation-free and data-free distillation method for accelerating diffusion and flow models while preserving strong generative capacity. Our key insight is to unify sample generation and score estimation with one student network. Thus, our framework uses only two models, a frozen teacher and a trainable student, and optimizes one objective. We prove that minimizing our objective indicates Wasserstein convergence of the student flow-map pushforwards to the teacher marginals. On ImageNet 256×256, our method attains an FID of 2.04 with a single function evaluation (1-NFE) and a 4-NFE FID of 1.37 within 40 epochs of training, surpassing the state-of-the-art distillation baselines without data. Our code code and model are available at https://consistentdmd.github.io/.
Yuxiang Fu, Qi Yan, Zike Wu +4
University of British Columbia · Vector Institute for AI · Canada CIFAR AI Chair
Score-based distillation methods train one-step diffusion models in two stages: they first train a teacher score model, then distill it into a one-step student model. However, this distillation process can introduce bias from two sources: errors in the teacher score model and student score estimate. We propose DiffRatio, a new approach for training one-step diffusion models without teacher supervision based on a density-ratio formulation that connects score-based distribution matching and GAN-style adversarial training. Instead of using a teacher score model to provide training targets, DiffRatio directly learns a log density ratio between the student and data distributions across diffusion time steps. As a pre-trained score model is only used to initialize the one-step generator and does not supervise training, this design simplifies the training pipeline, mitigates gradient estimation bias, and reduces the size of auxiliary networks. The learned density ratio can be used as a verifier, enabling a principled inference-time parallel scaling method that further improves sample quality without external rewards or extra sequential computation. DiffRatio achieves strong one-step generation results on CIFAR-10 and ImageNet (64x64 and 512x512), outperforming most teacher-supervised distillation methods.
Wenlin Chen, Mingtian Zhang, Jiajun He +4
1Bosch (China) Investment Ltd. · 2UCL · University of Cambridge +2
One-step distillation turns pretrained diffusion models into fast generators, but strong results often require substantial training machinery and compute. Pure adversarial distillation offers a simpler route, yet has struggled at modern text-to-image scale. To address this limitation, we propose ROAD (Representation-grounded One-step Adversarial Distillation), a one-stage framework that distills text-to-image models with one adversarial objective and a lightweight representation-space critic. First, we replace the classical convolutional head with a Lipschitz-controlled Transformer head with token-level text conditioning, which matches the semantic tokens of modern visual encoders. Second, we observed a fixed random feature mixing leads to long-horizon degradation. We attribute it to generator-critic co-adaptation around one realization-specific view, and periodically re-draw the mixing to keep the randomization effective. Third, for our dense and vector-valued logits, we introduce Jacobian R1 which penalizes the full input-logit Jacobian, overcoming pooled R1's limitation of neglecting most of logit sensitivities. A single adversarial objective then suffices: in one step, ROAD reaches 11.03 FID on SDXL and 13.73 on SD3.5-M on COCO-10K, surpassing the multi-step teachers at a fraction of distribution matching's training compute.
Xiaoou Li, Shan You, Tao Huang
Shanghai Jiao Tong University (sjtu.edu.cn) · SenseTime (sensetime.com)