acceptodds
Under review as a conference paper at ICLR 2027

ROAD: Modern Adversarial Training for Simple One-Step Distillation at Scale

Abstract

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 observe that fixed random feature mixing leads to long-horizon degradation. We attribute this 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 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.

Primary area:
generative models
Keywords:
one-step diffusion distillation, adversarial distillation, GAN, text-to-image generation, projected discriminator, Transformer critic, random feature mixing, Jacobian R1, diffusion models

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_YYJGOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

ROAD trains one-step SDXL and SD3.5-M generators with a single non-saturating GAN loss against 500K LAION-Aesthetics images, with the teacher used only as initialization. The critic follows the projected-GAN using a frozen DINOv3 ViT-B, with three innovations: Lipschitz-controlled ViT heads with token-level cross-attention, random feature mixing re-drawn every 5k steps against "generator–critic co-adaptation", and Jacobian R1, which penalizes the full input–logit Jacobian. Validation: one-step COCO-10K FID of 11.03 (SDXL) and 13.73 (SD3.5-M), better than the multi-step CFG teachers. CLIP and preference scores, ablations of each component, and a time-to-FID comparison showing that ROAD reaches ODE-initialized DMD2's best FID with 1/9.25 of the GPU-hours.

Critical strengths
  1. A very simple and effective training recipe. One loss, one stage, and no fake-score network, with a one-step performance comparable to the multi-step teacher model. This is a simple and cheaper method to one-step generators and would be appreciated by the community.

  2. Alignment is preserved without a teacher target. A pure real-data GAN keeps CLIP at 0.338 on SDXL (teacher 0.331). This addresses the known weakness of adversarial distillation of poor prompt adherence.

Critical weaknesses
  1. The FID comparison can be unfair The training target distribution is the LAION-Aesthetics data distribution (close to the coco dataset) (seen through DINOv3), so a lower COCO FID than a CFG teacher can be expected. In DMD2's own SDXL ablation, a GAN on the same 500K images reaches 13.77 FID (vs. 19.32) while CLIP drops to 0.307. On SD3.5-M, ROAD is worse than the teacher on HPSv2, ImageReward, PickScore and GenEval (Table 3).

  2. The GAN objective is mode-seeking, yet diversity is never measured. The diversity of the generator can be critical for further post-training or test time scaling.

Questions for authors
  1. Each generator step decodes latents and backpropagates through the VAE decoder and DINOv3. What share of time and memory does decoding take? Would a latent-space backbone such as the teacher's features (as in LADD) work in your framework?
  2. Are Figure 3 and Figure 4 saying the two sides of the same thing?
  3. What are the SDXL and SD3.5-M teachers' FID and CLIP under your protocol for CFG ∈ {1.5, 2, 3, 4.5, 6}? Where does ROAD sit relative to that curve?
  4. What are the FIDs of Tab. 1 rows 6–8 (Text conditioning)? Does the ViTGAN head in row 4 use token cross-attention?
6 comments
AuthorsOpenReview ↗

Response to Reviewer YYJG (1/3): Weakness 1

We sincerely thank you for your thorough and insightful review. We are grateful that you recognize "three innovations" in ROAD's critic, call ROAD "A very simple and effective training recipe" that "would be appreciated by the community", and note that it "addresses the known weakness of adversarial distillation of poor prompt adherence". Below, we respond to each of your points in turn; we hope these responses address your concerns, and we would be glad to clarify anything that remains unclear.

W1. The FID comparison can be unfair. The training target distribution is the LAION-Aesthetics data distribution (close to the coco dataset) (seen through DINOv3), so a lower COCO FID than a CFG teacher can be expected. In DMD2's own SDXL ablation, a GAN on the same 500K images reaches 13.77 FID (vs. 19.32) while CLIP drops to 0.307. On SD3.5-M, ROAD is worse than the teacher on HPSv2, ImageReward, PickScore and GenEval (Table 3).

We thank you for this concern. We believe that ROAD's FID gains over the teachers are fair, because ROAD's training images are not close to COCO, and the gains come from the ability of GANs to synthesize higher-fidelity images. Importantly, unlike the DMD2 ablation you cite, ROAD does not achieve its FID gains at the expense of text alignment, while maintaining competitive human-preference scores.

1. ROAD's FID gains over the teachers are fair.

  • On the data distribution, to the best of our knowledge, no study reports that the image distribution of LAION-Aesthetics is close to that of COCO, although both datasets consist of real images. We therefore evaluated their closeness directly. We computed the FID of 10,000 random images (mean of five draws) from our training dataset, which is constructed from LAION-Aesthetics. The FID is 46.7 against real COCO images, whereas ROAD's generated images achieve 11.25 FID against real COCO-10K images. This substantial difference suggests that ROAD does not obtain its low FID simply by imitating a training distribution already close to COCO.

  • On the source of the FID improvement, in our view, the compelling FID comes from the ability of GANs to synthesize higher-fidelity images. In the ablations of ADD and DMD2, removing the adversarial loss gives oversmoothed or oversaturated images with a much worse FID (ADD, Table 1d; DMD2, Table 4). Moreover, these ablations show that relying more on the adversarial objective can improve FID with the same real images. In your own example, which concerns four-step models, DMD2 with the GAN term alone reaches 13.77, against 19.32 for DMD2 with both score distillation and the GAN term (DMD2, Table 4). ROAD's purely adversarial objective makes the fullest use of this ability.

2. ROAD achieves a strong balance between FID and other quality metrics.

Importantly, ROAD overcomes the CLIP score degradation observed in the DMD2 ablation you cite. While that GAN-only model achieves a lower FID at the cost of a CLIP score of 0.307, ROAD achieves both a substantially lower FID and a CLIP score of 0.338. On SDXL, ROAD also surpasses its teacher on both human-preference metrics, HPSv2 and PickScore, as well as aesthetic score (Table 2). As for SD3.5-M, we agree that ROAD is behind the teacher on preference metrics. However, it is worth mentioning that, on SD3.5-M, there is no other one-step model to our knowledge, so we are competing with four-step baselines. ROAD, the first one-step method on SD3.5-M to our knowledge, surpasses the teacher on FID, CLIP score and aesthetic score, and its gaps on the preference metrics are small. At the teacher's best-FID guidance scale, ROAD is ahead on every reported metric (see Q3). These results suggest that ROAD achieves strong FID without a general degradation in other quality metrics.

AuthorsOpenReview ↗

Response to Reviewer YYJG (2/3): Weakness 2 and Question 1

W2. The GAN objective is mode-seeking, yet diversity is never measured. The diversity of the generator can be critical for further post-training or test time scaling.

Thank you for the insightful comment, and we agree that the concern is valid. We have now evaluated diversity. Experiments show that ROAD leads the distilled models on most diversity scores, and that it is more diverse than its teacher on all three main per-prompt scores.

Protocol. We follow DMD2 (App. B). Each method generates four images with seeds 0 to 3 for each of LADD's 128 PartiPrompts prompts. Per-prompt diversity is the mean pairwise distance among these images. The distances are LPIPS (DMD2), CLIP ViT-L/14 cosine distance (DP-DMD, ICML 2026), and the patch-level cosine distance of Harrington et al. (CVPR 2026). We compute the patch-level distance with PE-Spatial (Bolya et al., NeurIPS 2025) and DINOv3 in place of DINOv2. DINOv3 is for reference only, because ROAD's critic uses it. Recall (Kynkäänniemi et al., NeurIPS 2019) and coverage (Naeem et al., ICML 2020) use VGG-16 features with k=5k=5 on COCO-10K. DMDX has no public checkpoint, so we omit it. Higher is better. Bold marks the best distilled model, and underline marks the second-best. Teacher rows are not ranked.

│Per-prompt diversity\rlap{\textsf{Per-prompt diversity}}│Set-level diversity\rlap{\textsf{Set-level diversity}}
Method│LPIPSCLIPPE-SpatialDINOv3†\color{silver}{\textsf{DINOv3}^{\dagger}}│RecallCoverage
SDXL││
SDXL base (cfg 6)│0.6330.1680.5350.317\color{silver}0.317│0.6410.665
Hyper-SDXL│0.5860.0900.4440.137\color{silver}0.137│0.643‾\underline{\textsf{0.643}}0.311
DMD2│0.6140.1270.4960.270\color{silver}0.270│0.5700.766‾\underline{\textsf{0.766}}
SiDa│0.671‾\underline{\textsf{0.671}}0.1840.554‾\underline{\textsf{0.554}}0.379‾\color{silver}{\underline{0.379}}│0.5720.767
SDXL-Turbo│0.6050.0970.4850.194\color{silver}0.194│0.5510.741
SDXL-Lightning│0.6270.1670.5010.289\color{silver}0.289│0.5720.639
NitroSD-Realism│0.6390.1150.5010.243\color{silver}0.243│0.6260.687
ROAD (ours)│0.6740.180‾\underline{\textsf{0.180}}0.5560.402\color{silver}{\mathbf{0.402}}│0.6610.705
SD3.5-M││
SD3.5-M (cfg 4.5)│0.6100.1280.4820.274\color{silver}0.274│0.7210.670
SD3.5-M-Turbo│0.5650.150‾\underline{\textsf{0.150}}0.4420.293‾\color{silver}{\underline{0.293}}│0.5610.611
SenseFlow│0.6410.1210.4910.216\color{silver}0.216│0.6180.411
SwD│0.5210.0930.4040.173\color{silver}0.173│0.6750.477
SiD-DiT│0.643‾\underline{\textsf{0.643}}0.1360.498‾\underline{\textsf{0.498}}0.271\color{silver}0.271│0.6030.686‾\underline{\textsf{0.686}}
ROAD (ours)│0.6500.1570.5360.337\color{silver}{\mathbf{0.337}}│0.646‾\underline{\textsf{0.646}}0.803

Results. The five main scores on SDXL and SD3.5-M give ten comparisons. Among the distilled models, ROAD has the highest value in 7 of the 10 comparisons. It has the second-highest value in 2 further comparisons. No other method has the highest value in more than 2 comparisons. The exception is SDXL coverage, where ROAD ranks fourth after SiDa, DMD2 and SDXL-Turbo. On both models, ROAD is also more diverse than its teacher on all three per-prompt scores.

Q1. What share of time and memory does decoding take? Would a latent-space backbone such as the teacher's features (as in LADD) work in your framework?

ROAD's reported cost already includes decoding, and we believe that its critic can use a latent-space backbone, because its components do not depend on the backbone.

On decoding, its cost is already included in the training GPU-hours of Fig. 1. Per generator step, VAE decoding takes 21.61% of the time and 6.69% of the peak memory.

On a latent-space backbone, ROAD's head, refresh and JR1 act on any grid of feature tokens, so the backbone can be exchanged. ROAD already works with three different frozen backbones, PE-Spatial, SigLIP2 and DINOv3, with close FID (Table 4b, L478–482). The teacher's features, as in LADD, fit ROAD naturally when the teacher is a DiT whose tokens form one uniform grid, as LADD shows with SD3. For a UNet teacher such as SDXL, the features form a multi-resolution pyramid, so ROAD will have different token numbers for different levels. In this latent-space setting, the real samples will be either generated by teachers as LADD or obtained by the VAE encoder. We thank you for this insightful question, and we think that extending ROAD to a latent-space critic is worth future work.

AuthorsOpenReview ↗

Response to Reviewer YYJG (3/3): Questions 2–4

Q2. Are Figure 3 and Figure 4 saying the two sides of the same thing?

Yes, they are the two arms of one controlled comparison. The runs share seeds, data and the diagnostic set, and they differ only in whether the mixing is fixed or re-drawn every 5k steps.

Q3. What are the SDXL and SD3.5-M teachers' FID and CLIP under your protocol for CFG ∈ {1.5, 2, 3, 4.5, 6}? Where does ROAD sit relative to that curve?

Thank you for this question. Our original teacher baselines use the established CFG settings of 6 for SDXL and 4.5 for SD3.5-M, consistent with official model guidance and previous studies such as DMD2. Following your suggestion, we have conducted these experiments, and the following tables give the results. Bold marks the best value and underlining the second-best value in each column.

Each table has two ROAD rows, and they differ only in the data used to select the checkpoint. The row "ROAD, 1 step (val set)" is the checkpoint of the submitted paper, which we selected on COCO-10K. This selection follows our closest predecessors, DMD2 and Projected GAN, and established GAN papers such as StyleGAN2 and StyleGAN3. To address a concern of Reviewer aA4p, we drew 10,000 COCO-2014 validation images and captions at random, disjoint from DMD2's COCO-10K, and we selected the checkpoint on this disjoint set. The row "ROAD, 1 step (disjoint set)" reports this checkpoint, which the revised Tables 2 and 3 now use. We show the val-set rows in gray for reference, and we do not rank them. Kindly refer to our response to Reviewer aA4p (W1 and Q1) for details on the checkpoint selection issue.

SDXL base, cfgFID ↓CLIP ↑HPSv2 ↑Pick ↑Aes ↑
1.533.570.27525.7120.696.08
224.070.29626.4621.166.13
319.140.31527.3221.656.16
4.518.49‾\underline{\textsf{18.49}}0.32627.9721.966.18
6 (default)19.290.331‾\underline{\textsf{0.331}}28.33‾\underline{\textsf{28.33}}22.11‾\underline{\textsf{22.11}}6.19‾\underline{\textsf{6.19}}
ROAD, 1 step (val set)\color{gray}{\textsf{ROAD, 1 step (val set)}}11.03\color{gray}{11.03}0.338\color{gray}{0.338}28.58\color{gray}{28.58}22.13\color{gray}{22.13}6.67\color{gray}{6.67}
ROAD, 1 step (disjoint set)11.250.33828.6122.146.62
SD3.5-M, cfgFID ↓CLIP ↑HPSv2 ↑IR ↑Pick ↑Aes ↑GE ↑
1.515.19‾\underline{\textsf{15.19}}0.32527.420.3921.295.850.52
215.460.33528.060.6721.735.970.61
316.230.34128.700.8822.136.060.65
4.5 (default)18.010.34129.191.0022.386.12‾\underline{\textsf{6.12}}0.69
619.120.34429.09‾\underline{\textsf{29.09}}0.94‾\underline{\textsf{0.94}}22.29‾\underline{\textsf{22.29}}6.090.66‾\underline{\textsf{0.66}}
ROAD, 1 step (val set)\color{gray}{\textsf{ROAD, 1 step (val set)}}13.73\color{gray}{13.73}0.342\color{gray}{0.342}28.85\color{gray}{28.85}0.94\color{gray}{0.94}22.00\color{gray}{22.00}6.15\color{gray}{6.15}0.66\color{gray}{0.66}
ROAD, 1 step (disjoint set)13.890.343‾\underline{\textsf{0.343}}28.900.9321.856.170.66

On SDXL, ROAD is better than the teacher at every guidance scale on all five metrics. Even the teacher's best value in each column, such as FID 18.49 at cfg 4.5 or CLIP 0.331 at cfg 6, is behind ROAD. This shows that no guidance scale lets the teacher reach ROAD.

On SD3.5-M, ROAD has the best FID and aesthetic score, and it is better than the teacher's best-FID setting (i.e., cfg 1.5) on every metric. A higher guidance scale raises the teacher's preference scores but worsens its FID, so that at cfg 4.5 and 6 the teacher leads ROAD on most preference metrics while its FID is more than 4 points worse. These conclusions also hold for both ROAD checkpoints.

Q4. What are the FIDs of Tab. 1 rows 6–8 (Text conditioning)? Does the ViTGAN head in row 4 use token cross-attention?

The table below gives the FID and CLIP score of Table 1, rows 6–8, with row 4 for reference.

Table 1 rowCritic headText conditioningFID ↓CLIP ↑
4ViTGANToken-level cross-attention15.400.319
6ViT (ours)Unconditional14.310.320
7ViT (ours)Miyato-style pooled-text projection14.270.322
8ViT (ours)Token-level cross-attention14.540.331

Rows 1 and 2 use Miyato-style pooled-text projection, and row 4 (ViTGAN) uses token-level cross-attention. We note that ViTGAN improves FID significantly over the convolutional head but improves CLIP score only marginally. We attribute this to its sub-optimal Lipschitz control strategies. Namely, it uses L2 attention with tied query and key projections, and applies spectral normalization to linear weights only at initialization instead of throughout training.

Reviewer_AqeXOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ROAD, a one-stage adversarial distillation framework for one-step text-to-image generation. The method uses a lightweight representation-space critic with a Lipschitz-controlled Transformer head and token-level text conditioning, periodically re-samples random feature mixing to reduce generator–critic co-adaptation, and introduces a Jacobian R1 regularizer for dense vector-valued logits. With only a single adversarial objective, ROAD achieves strong one-step generation performance on SDXL and SD3.5-M, outperforming their multi-step teachers on COCO-10K while requiring substantially less training compute than distribution-matching-based distillation methods.

Critical strengths
  1. The training pipeline is simple. A one-step generator can be obtained using only single-stage adversarial training, without continuously querying the teacher or training an additional large scoring network.
  2. The method achieves a good balance between performance and efficiency. ROAD obtains low FID on both SDXL and SD3.5-M, while also demonstrating a clear training-compute advantage over DMD2.
Critical weaknesses
  1. The method still relies on multiple stabilization techniques. Although the training objective itself is simple, the final performance depends on several components, including spectral normalization, cosine attention, residual scaling, periodic mixing refresh, and Jacobian regularization. As a result, the overall training recipe remains relatively complex.
  2. The mechanism behind random mixing refresh is not sufficiently explained. Periodically refreshing the random mixing helps alleviate training degradation, but the theoretical analysis is based on simplified assumptions and does not guarantee the stability of the full adversarial training process. In addition, the paper does not provide a clear principle for how the refresh frequency should scale with model or dataset size.
  3. The evaluated base models are relatively dated. The experiments focus mainly on SDXL and SD3.5-M, with no evaluation on newer-generation text-to-image models (e.g., FLUX, Z-Image, Qwen-Image). It is therefore unclear whether the method can scale to stronger generators while preserving the same quality and efficiency advantages.
Questions for authors

Refer to weakness.

6 comments
AuthorsOpenReview ↗

Response to Reviewer AqeX (1/2): Weaknesses 1 and 2

We sincerely thank you for your constructive review. We are grateful that you find that "The training pipeline is simple", that ROAD "achieves a good balance between performance and efficiency", and that it "achieves strong one-step generation performance on SDXL and SD3.5-M". Below, we respond to each of your points in turn; we hope these responses address your concerns, and we would be glad to clarify anything that remains unclear.

W1. The method still relies on multiple stabilization techniques. Although the training objective itself is simple, the final performance depends on several components, including spectral normalization, cosine attention, residual scaling, periodic mixing refresh, and Jacobian regularization. As a result, the overall training recipe remains relatively complex.

Thank you for this concern, and we respectfully disagree. We believe that, compared with other adversarial distillation methods, the five components you list do not make ROAD's training recipe more complex.

Above all, we distinguish the complexity of the training pipeline from the complexity of the critic. In its training pipeline, ROAD uses one GAN objective in one stage, whereas most of the competing methods combine two to four objectives over up to six stages. We think that the training pipeline matters more, because it decides how many objectives must be balanced and how many stages must be run.

In fact, the critic is not more complex, because all five components belong to it and form three parts, namely the head architecture, the random mixing and the regularization.

  1. The head architecture adds no tuning. Every adversarial method must design a critic head. Compared with these heads, our spectral normalization, cosine attention and residual scaling add no hyperparameter to search, because spectral normalization has none and the attention temperatures and residual scales are learned (Sec. E.1).
  2. Mixing refresh adds one simple setting. Projected-GAN critics, such as those of Projected GAN and StyleGAN-XL, already draw a random mixing once. Compared with them, refresh only re-draws it periodically, with the same interval on both models and no extra loss, network or stage (Table 6).
  3. The regularization adds no new hyperparameter. Adversarial methods already regularize the critic. For example, ADD uses R1 and D2O uses pooled R1. Compared with these penalties, JR1 keeps one weight, as R1 has, and adds no loss term or network.

W2. The mechanism behind random mixing refresh is not sufficiently explained. Periodically refreshing the random mixing helps alleviate training degradation, but the theoretical analysis is based on simplified assumptions and does not guarantee the stability of the full adversarial training process. In addition, the paper does not provide a clear principle for how the refresh frequency should scale with model or dataset size.

Thank you for this comment. Below, we clarify the mechanism of refresh, the stability of ROAD's full training, and the role of the refresh interval.

1. On the mechanism, the paper explains refresh with a chain of analyses. Augmentation first rules out ordinary overfitting of the critic (L255–257, Table 8). PFUC and the logit margin then reveal a persistent critic state under a fixed draw (L262–287, Fig. 4), and fresh-draw tests trace this state to co-adaptation with the one draw (L289–301). The linear surrogate explains why such co-adaptation leaves weakly measured directions uncorrected (L303–315 and L320–322, Prop. B.1), and re-drawing removes the persistent state together with the degradation (L316–344, Fig. 5).

2. On stability, as far as we know, no modern deep GAN has a guarantee of stable full training without simplifications or strict conditions. The existing guarantees either assume simplified games, such as bilinear games (Daskalakis et al., 2018), or cover only the end of training, when the generator and the critic are already close to the solution (Mescheder et al., 2018). Notably, our linear surrogate is a simplified analysis that serves to explain the co-adaptation mechanism, and it makes no claim on stability. ROAD's stability comes from Lipschitz control, the refresh mechanism and JR1, because the head collapses without Lipschitz control (Table 1), and training improves steadily with refresh (Fig. 6) and with JR1 (Fig. 7).

3. On the interval, a precise scaling rule is not needed, because the results are insensitive to the interval over a wide range. Intervals from 2k to 10k steps give similar FID (Fig. 6), and the interval of 5k steps transferred from a UNet to an MMDiT without retuning (Table 6). We expect the same on larger datasets, because the interval depends on how fast the critic adapts to a draw.

AuthorsOpenReview ↗

Response to Reviewer AqeX (2/2): Weakness 3

W3. The evaluated base models are relatively dated. The experiments focus mainly on SDXL and SD3.5-M, with no evaluation on newer-generation text-to-image models (e.g., FLUX, Z-Image, Qwen-Image). It is therefore unclear whether the method can scale to stronger generators while preserving the same quality and efficiency advantages.

Thank you for this comment. We believe that our base models remain the standard benchmarks of current distillation research. We also appreciate your suggestion to evaluate newer models, and have conducted an additional experiment on FLUX.2 [klein] 4B Base to address this concern.

1. Standard Benchmarks. SDXL and SD3.5-M remain the main benchmarks of distillation papers at top venues in 2026. For example, SenseFlow and SwD (ICLR 2026) both distill SDXL and SD3.5, Flash-DMD (CVPR 2026) distills SDXL and SD3-Medium, and DP-DMD (ICML 2026) distills SD3.5-M. Moreover, the baselines of Tables 2 and 3 are public checkpoints for these two models, except DMDX, whose values come from its paper, so every other method is evaluated under one protocol (L1418–1420). This shows that the two models give the standard and fairest comparison of the field.

2. Validation on a Newer Model. Following your suggestion, we extended ROAD to FLUX.2 [klein] 4B Base, a newer 4-billion-parameter flow transformer, and distilled it into a one-step generator. After only 6k training iterations, ROAD achieves an FID of 17.42, substantially better than the multi-step teacher's 23.26, while also improving CLIP score (0.340 vs. 0.336). ROAD also slightly surpasses the teacher on HPSv2 and aesthetic score, matches its ImageReward, and remains slightly behind on PickScore. These results suggest that ROAD generalizes to newer and larger architectures while maintaining strong generation quality. We have added this experiment to Sec. E.4 (Table 9) of the revised paper.

MethodNFEFID ↓CLIP ↑HPSv2 ↑IR ↑Pick ↑Aes ↑
FLUX.2 [klein] 4B Basemulti23.260.33629.130.9622.566.12
ROAD (ours, 6k iterations)117.420.34029.210.9622.006.16

3. Scalability. ROAD has now demonstrated strong results from SD3.5-M (2.5B) to FLUX.2 (4B). As you note in your strengths, ROAD trains "without continuously querying the teacher or training an additional large scoring network". Its trainable critic heads are lightweight and sit on a frozen DINOv3 ViT-B, whatever the size of the generator (Table 1). In contrast, DMD2 runs the teacher during training and trains a score network as large as the generator. This suggests that ROAD's training-compute advantage over DMD2 could widen for larger generators. Although the limited rebuttal period and computational resources prevent us from evaluating even larger models, these encouraging results suggest that ROAD could scale further to models such as Z-Image and Qwen-Image, which we leave for future work.

Dear Reviewer AqeX,

Thank you very much for your encouraging feedback and valuable suggestions regarding our work.

As we approach the end of the discussion period, we wanted to briefly follow up and ask whether our responses have sufficiently clarified the questions you raised.

If there are any outstanding concerns or additional points you would like us to discuss, we would be more than happy to address them.

We truly appreciate your time and effort in reviewing our paper.

Best regards,

The Authors

Reviewer_aA4pOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes a adversarial method for distilling text-to-image diffusion models (SDXL, SD3.5-M) into one-step generators. The student is initialized from the teacher and trained adversarially against 500K real LAION images with a strong discriminator. The paper reports a COCO-10K FID of 11.03 on SDXL and 13.73 on SD3.5-M which are the best FIDs among the compared baselines, and the SDXL result reaches DMD2’s best FID at about 1/9 of its GPU-hours.

Critical strengths
  • A single-stage, pure-adversarial recipe with no online score network, ODE-pair regression, or reward model is practically valuable. The compute comparison in Fig. 1 is a meaningful result.
  • The paper covers a very throrough ablation: the head design, text conditioning, backbone choice, regularizer type and coefficient, refresh interval, and mixing kernels. These can be the foundation for future model to base on.
Critical weaknesses
  • If I understand correctly, based on Appendix E.2 section at L1359, the authors select the checkpoint based on the same validation COCO-10K subset used for final evaluation which is clearly not fair with other baselines(?).
  • The headline claims rest on FID alone. On SDXL, ROAD is not better on human-preference metrics. On SD3.5-M, ROAD has among the lowest HPSv2, ImageReward, and PickScore scores.
  • Diversity and possible metric are not assessed. One-step adversarial generators are prone to reduced diversity, yet the paper does not report anything about diversity.
  • Methodological novelty is modest, and the theory adds little. There are no new insight here, mostly improve existing architectures and adopt them into modern text-to-image diffusion model.
Questions for authors
  • Can you report the number for ROAD and compare it against baseline when selecting the checkpoint based on a disjoint COCO-10K set
  • What are the teacher’s FID values at lower guidance scales? Does ROAD still surpass the teacher at its best-FID guidance?
  • Can you report recall/coverage and per-prompt diversity (e.g., LPIPS, DINOv3, CLIP diversity across seeds) for ROAD and the baselines to evaluate their diversity?
8 comments
AuthorsOpenReview ↗

Response to Reviewer aA4p (1/3): Weakness 1 and Question 1

We sincerely thank you for your careful reading and your thoughtful comments. We are grateful that you find the recipe "practically valuable", the compute comparison in Fig. 1 "a meaningful result", and the ablation "very thorough", one that "can be the foundation for future model to base on". Below, we respond to each of your points in turn; we hope these responses address your concerns, and we would be glad to clarify anything that remains unclear.

W1. If I understand correctly, based on Appendix E.2 section at L1359, the authors select the checkpoint based on the same validation COCO-10K subset used for final evaluation which is clearly not fair with other baselines(?).

Q1. Can you report the number for ROAD and compare it against baseline when selecting the checkpoint based on a disjoint COCO-10K set

Thank you for the careful reading of Sec. E.2, and your understanding of ROAD's checkpoint selection is correct.

Notably, our checkpoint selection follows the closest predecessors and the established GAN papers. DMD2, whose COCO-10K protocol we adopt, uses COCO-10K both to choose its design, by evaluating its ablation variants on it (DMD2, Table 4), and to report its final results (DMD2, Table 2). During training, its released code also evaluates every saved checkpoint on COCO-10K. Projected GAN trains its own model and its competing GAN methods, and reports for each of them the checkpoint with the lowest FID during training (Sauer et al., NeurIPS 2021, Sec. 5). This FID compares 50,000 generated images with the full training dataset, so one set both selects the checkpoint and produces the reported FID. StyleGAN2 and StyleGAN3 likewise report, for each trained model, the checkpoint with the lowest FID on the set that also produces the reported FID (Karras et al., CVPR 2020, Table 1; Karras et al., NeurIPS 2021, Fig. 12). This shows that, in the closest predecessors and in the established GAN papers, the set that produces the reported FID is also the set used to select checkpoints or design choices.

Nevertheless, we think your concern makes sense. To address it, we conducted an experiment using 10,000 COCO-2014 validation images and captions drawn at random and disjoint from DMD2's COCO-10K.

ROAD's lead holds when its checkpoint is selected on the disjoint set, as you request in Q1. We drew the disjoint set with DMD2's rule of one caption per image and a seed fixed before any evaluation. Among the same checkpoints saved every 500 training steps, we selected the one with the lowest FID on this disjoint set and evaluated it on COCO-10K. Its COCO-10K FID is 11.25 on SDXL, against 11.03 in the submitted Table 2, and 13.89 on SD3.5-M, against 13.73 in the submitted Table 3. The tables below provide a detailed comparison of the SDXL and SD3.5-M results, with all metrics evaluated on the original COCO-10K.

SDXLFID ↓CLIP ↑HPSv2 ↑Pick ↑Aes ↑
SDXL base (cfg 6)19.290.33128.3322.116.19
ROAD (val set)11.030.33828.5822.136.67
ROAD (disjoint set)11.250.33828.6122.146.62
SD3.5-MFID ↓CLIP ↑HPSv2 ↑IR ↑Pick ↑Aes ↑GE ↑
SD3.5-M (cfg 4.5)18.010.34129.191.0022.386.120.69
ROAD (val set)13.730.34228.850.9422.006.150.66
ROAD (disjoint set)13.890.34328.900.9321.856.170.66

For ROAD, "val set" and "disjoint set" denote the datasets used for checkpoint selection; all reported metrics are evaluated on the original COCO-10K.

This shows that selecting on COCO-10K changes ROAD's FID by at most 0.22, which is far less than ROAD's lead over the next method in Tables 2 and 3. On SD3.5-M, the checkpoint selected on the disjoint set still has a lower FID than the teacher by 4.12 points. We have updated the ROAD rows of Tables 2 and 3 with the results of the disjoint-set checkpoint, and we have described the disjoint selection in Sec. E.2 (L1413–1419) of the revised paper.

AuthorsOpenReview ↗

Response to Reviewer aA4p (2/3): Weaknesses 2 and 3, and Question 3

W2. The headline claims rest on FID alone. On SDXL, ROAD is not better on human-preference metrics. On SD3.5-M, ROAD has among the lowest HPSv2, ImageReward, and PickScore scores.

Thank you for this concern. We agree that, on SD3.5-M, ROAD is behind the teacher on HPSv2, ImageReward and PickScore. We have therefore revised the headline claims at the end of the abstract (L24–29) and at the end of Sec. 1 (L121–128). Kindly refer to these passages in the revised paper. Nevertheless, it is worth mentioning that, on SD3.5-M, there is no 1-step model except ours, so we are competing with 4-step baselines. ROAD, the first one-step model on SD3.5-M to our knowledge, surpasses the teacher on FID, CLIP score and aesthetic score, and at the teacher's best-FID guidance scale it is ahead on every metric (kindly see our response to Q2 below).

On SDXL, ROAD is slightly better than its teacher on both human-preference metrics, i.e., ROAD vs teacher: 28.61 vs 28.33 (HPSv2) and 22.14 vs 22.11 (Pick). Its HPSv2 and PickScore exceed the teacher's at every guidance scale, and its aesthetic score is the highest of all methods; kindly see Table 2 and our response to Q2 below. Among the distilled models, only Hyper-SDXL and NitroSD-Realism achieve higher HPSv2 scores. However, Hyper-SDXL requires 4 objectives and 6 stages, including explicit human-preference reward learning, while NitroSD-Realism uses 2 objectives and 3 stages. In contrast, ROAD achieves competitive preference scores with a single objective and a single training stage. This shows that ROAD's FID gain does not cost preference on SDXL.

W3. Diversity and possible metric are not assessed. One-step adversarial generators are prone to reduced diversity, yet the paper does not report anything about diversity.

Q3. Can you report recall/coverage and per-prompt diversity (e.g., LPIPS, DINOv3, CLIP diversity across seeds) for ROAD and the baselines to evaluate their diversity?

Thank you for the insightful comment and the question. Following your suggestion, we have evaluated per-prompt diversity and set-level diversity.

Protocol. We follow DMD2 (App. B). Each method generates four images with seeds 0 to 3 for each of LADD's 128 PartiPrompts prompts. Per-prompt diversity is the mean pairwise distance among these images. The distances are LPIPS (DMD2), CLIP ViT-L/14 cosine distance (DP-DMD), and the patch-level cosine distance of Harrington et al. We compute the patch-level distance with PE-Spatial and DINOv3 in place of DINOv2. DINOv3 is for reference only, because ROAD's critic uses it. Recall and coverage use VGG-16 features with k=5k=5 on COCO-10K. DMDX has no public checkpoint, so we omit it. Higher is better. Bold marks the best distilled model, and underline marks the second-best. Teacher rows are not ranked.

│Per-prompt diversity\rlap{\textsf{Per-prompt diversity}}│Set-level diversity\rlap{\textsf{Set-level diversity}}
Method│LPIPSCLIPPE-SpatialDINOv3†^{\dagger}│RecallCoverage
SDXL││
SDXL base (cfg 6)│0.6330.1680.5350.317│0.6410.665
Hyper-SDXL│0.5860.0900.4440.137│0.643‾\underline{\textsf{0.643}}0.311
DMD2│0.6140.1270.4960.270│0.5700.766‾\underline{\textsf{0.766}}
SiDa│0.671‾\underline{\textsf{0.671}}0.1840.554‾\underline{\textsf{0.554}}0.379‾\underline{\textsf{0.379}}│0.5720.767
SDXL-Turbo│0.6050.0970.4850.194│0.5510.741
SDXL-Lightning│0.6270.1670.5010.289│0.5720.639
NitroSD-Realism│0.6390.1150.5010.243│0.6260.687
ROAD (ours)│0.6740.180‾\underline{\textsf{0.180}}0.5560.402│0.6610.705
SD3.5-M││
SD3.5-M (cfg 4.5)│0.6100.1280.4820.274│0.7210.670
SD3.5-M-Turbo│0.5650.150‾\underline{\textsf{0.150}}0.4420.293‾\underline{\textsf{0.293}}│0.5610.611
SenseFlow│0.6410.1210.4910.216│0.6180.411
SwD│0.5210.0930.4040.173│0.6750.477
SiD-DiT│0.643‾\underline{\textsf{0.643}}0.1360.498‾\underline{\textsf{0.498}}0.271│0.6030.686‾\underline{\textsf{0.686}}
ROAD (ours)│0.6500.1570.5360.337│0.646‾\underline{\textsf{0.646}}0.803

Results. The five main scores on SDXL and SD3.5-M give ten comparisons. Among the distilled models, ROAD has the highest value in 7 of the 10 comparisons. It has the second-highest value in 2 further comparisons. The exception is SDXL coverage, where ROAD ranks fourth after SiDa, DMD2 and SDXL-Turbo. On both models, ROAD is also more diverse than its teacher on all three per-prompt scores.

AuthorsOpenReview ↗

Response to Reviewer aA4p (3/3): Weakness 4 and Question 2

W4. Methodological novelty is modest, and the theory adds little. There are no new insight here, mostly improve existing architectures and adopt them into modern text-to-image diffusion model.

We agree that our ROAD builds on the existing projected GAN, which is an active and influential research line of works including Projected GAN (NeurIPS 2021), StyleGAN-XL (SIGGRAPH 2022), StyleGAN-T (ICML 2023), ADD (ECCV 2024) and D2O (ICML 2025).

However, our modernization of projected GANs introduces new findings and three substantive contributions.

  1. We find that convolutional heads underexploit modern representations in the projected GAN framework. We propose Lipschitz-controlled Transformer heads with token-level text conditioning and further derive a theoretical Lipschitz bound for them. To the best of our knowledge, this is the first use of such heads in this framework. Our Transformer heads substantially improve both FID and CLIP scores over the existing heads.
  2. We reveal the co-adaptation of the generator and the critic under fixed random mixing. We analyze its cause in depth through the gains of the random mixing matrix and propose a refresh mechanism that resolves this limitation. Moreover, in Sec. B.2 of the appendix, we build a simplified linear surrogate of the GAN for a theoretical analysis of this co-adaptation.
  3. We identify the limitation of classical pooled R1. We propose JR1 to exploit the full sensitivity of the vector-valued logits. In Sec. C.1 of the appendix, we prove that JR1 removes two limitations of pooled R1: pooled R1 (i) leaves n−1n-1 of the nn directions of output logits unpenalized and (ii) can incur zero penalty for a nonzero Jacobian.

Together, these contributions enable strong one-step generation with a single adversarial objective and one training stage. We believe that ROAD, a simpler and cheaper method, would be appreciated by the community (Reviewer YYJG) and, as you note, can be the foundation for future models to build on. Therefore, we would sincerely appreciate it if you would reconsider your evaluation of the novelty and contributions of ROAD.

Q2. What are the teacher's FID values at lower guidance scales? Does ROAD still surpass the teacher at its best-FID guidance?

Thank you for this insightful suggestion. Our original teacher baselines use the established CFG settings of 6 for SDXL and 4.5 for SD3.5-M, consistent with official model guidance and previous studies such as DMD2. We have now evaluated the teacher at five guidance scales, from 1.5 to 6, on both SDXL and SD3.5-M.

The results are given in the following tables, where bold marks the best and underlining the second-best value in each column. The gray rows are for reference and are not ranked.

SDXL base, cfgFID ↓CLIP ↑HPSv2 ↑Pick ↑Aes ↑
1.533.570.27525.7120.696.08
224.070.29626.4621.166.13
319.140.31527.3221.656.16
4.518.49‾\underline{\textsf{18.49}}0.32627.9721.966.18
6 (default)19.290.331‾\underline{\textsf{0.331}}28.33‾\underline{\textsf{28.33}}22.11‾\underline{\textsf{22.11}}6.19‾\underline{\textsf{6.19}}
ROAD, 1 step (val set)\color{gray}{\textsf{ROAD, 1 step (val set)}}11.03\color{gray}{11.03}0.338\color{gray}{0.338}28.58\color{gray}{28.58}22.13\color{gray}{22.13}6.67\color{gray}{6.67}
ROAD, 1 step (disjoint set)11.250.33828.6122.146.62
SD3.5-M, cfgFID ↓CLIP ↑HPSv2 ↑IR ↑Pick ↑Aes ↑GE ↑
1.515.19‾\underline{\textsf{15.19}}0.32527.420.3921.295.850.52
215.460.33528.060.6721.735.970.61
316.230.34128.700.8822.136.060.65
4.5 (default)18.010.34129.191.0022.386.12‾\underline{\textsf{6.12}}0.69
619.120.34429.09‾\underline{\textsf{29.09}}0.94‾\underline{\textsf{0.94}}22.29‾\underline{\textsf{22.29}}6.090.66‾\underline{\textsf{0.66}}
ROAD, 1 step (val set)\color{gray}{\textsf{ROAD, 1 step (val set)}}13.73\color{gray}{13.73}0.342\color{gray}{0.342}28.85\color{gray}{28.85}0.94\color{gray}{0.94}22.00\color{gray}{22.00}6.15\color{gray}{6.15}0.66\color{gray}{0.66}
ROAD, 1 step (disjoint set)13.890.343‾\underline{\textsf{0.343}}28.900.9321.856.170.66

On SDXL, ROAD is better than the teacher at every guidance scale on all five metrics. Even the teacher's best value in each column, such as FID 18.49 at cfg 4.5 or CLIP 0.331 at cfg 6, is behind ROAD. This shows that no guidance scale lets the teacher reach ROAD.

On SD3.5-M, ROAD has the best FID and aesthetic score, and it is better than the teacher's best-FID setting (i.e., cfg 1.5) on every metric. A higher guidance scale raises the teacher's preference scores but worsens its FID, so that at cfg 4.5 and 6 the teacher leads ROAD on most preference metrics while its FID is more than 4 points worse. These conclusions also hold for both ROAD checkpoints.

Comments on the paper

2 comments
AuthorsOpenReview ↗

"ROAD (Submission 15585): general response to the AC and all reviewers"

We thank the AC for handling our paper, and we thank the three reviewers for their careful reviews and constructive suggestions. We have addressed every point in the individual responses.

Strengths identified by the reviewers

  • Training recipe. YYJG calls ROAD "a very simple and effective training recipe" that "would be appreciated by the community". AqeX writes that "the training pipeline is simple". aA4p finds the recipe "practically valuable".
  • Technical contributions. YYJG recognizes "three innovations" in the critic. YYJG also notes that ROAD "addresses the known weakness of adversarial distillation of poor prompt adherence".
  • Results and efficiency. AqeX writes that ROAD "achieves strong one-step generation performance on SDXL and SD3.5-M" and "achieves a good balance between performance and efficiency". aA4p calls the compute comparison of Fig. 1 "a meaningful result".

Main point of the paper

We propose a one-step distillation method that uses a single adversarial objective in a single training stage, by modernizing the projected-GAN line of research. Our modernization introduces new findings and three substantive contributions.

We find that convolutional heads underexploit modern representations, and we therefore propose Lipschitz-controlled Transformer heads with token-level text conditioning. We also reveal that the generator and the critic co-adapt under a fixed random mixing. To break this co-adaptation, we propose a periodic refresh of the mixing. In addition, we identify the limitations of the classical pooled R1, and we address them with Jacobian R1 (JR1).

We evaluate ROAD on the UNet-based SDXL and on the MMDiT-based SD3.5-M, and ROAD achieves compelling results in FID and CLIP score. ROAD also converges much more quickly, 9.25 times faster than DMD2 on SDXL.

Common concerns (Reviewers YYJG & aA4p)

  • Diversity against the competing methods. YYJG notes that the paper does not evaluate diversity. aA4p asks for recall, coverage, and per-prompt diversity. Our evaluation on both models shows that ROAD is the most consistently diverse distilled model.
  • ROAD versus the teachers across CFG scales. YYJG and aA4p ask whether ROAD still surpasses the teachers at lower classifier-free guidance (CFG) scales. We evaluated both teachers at CFG scales from 1.5 to 6. ROAD surpasses the teacher on all metrics at every CFG scale on SDXL, and at the teacher's best-FID scale on SD3.5-M.
  • Performance on SD3.5-M. YYJG and aA4p note that ROAD trails the SD3.5-M teacher on human-preference scores. We agree that, at its default CFG scale, the teacher leads ROAD on these scores by small margins. ROAD still surpasses the teacher in FID, CLIP score, and aesthetic score. Moreover, ROAD is a one-step generator, whereas the competing SD3.5-M baselines use four sampling steps.

Major concerns

  • Checkpoint selection, novelty, and insight (Reviewer aA4p). aA4p points out an issue with ROAD's checkpoint selection. We now select the checkpoint on a held-out set that is disjoint from COCO-10K. ROAD's FID changes by at most 0.22, and ROAD still has the lowest FID among all methods and the teachers. aA4p also thinks that the methodological novelty is modest, and notes a lack of insight. We now state our findings and the key contributions above explicitly in the paper.
  • FID fairness and the training distribution (Reviewer YYJG). YYJG suggests that our LAION-Aesthetics training images may be close to COCO, so a low FID on COCO would be expected. We computed the FID between our training images and COCO, and it is 46.7, which is far above ROAD's 11.25. ROAD's low FID therefore does not come from a training distribution that is already close to COCO.
  • Training recipe, refresh mechanism, and scalability (Reviewer AqeX). AqeX finds the training recipe complex, and also asks for a clearer mechanism of refresh and a guarantee of global stability. All the components that AqeX points out belong to the critic, and together they add only one hyperparameter, the refresh interval. Like other modern deep GANs, ROAD has no theoretical guarantee of global stability. In short, refresh prevents the generator and the critic from co-adapting to one fixed mixing. Finally, AqeX asks whether ROAD scales to newer models such as FLUX. In a new experiment on FLUX.2 [klein] 4B, one-step ROAD achieves an FID of 17.42 after only 6k training iterations, compared with 23.26 for the multi-step teacher.

Summary of revisions

Revisions are highlighted in light blue.

LocationReviewers
AbstractYYJG, aA4p
Sec. 1YYJG, aA4p
Fig. 1Presentation
Sec. 3.1, Table 1YYJG
Tables 2 and 3aA4p
Sec. E.2aA4p
Sec. E.4, Table 8AqeX
Sec. E.4, Table 9YYJG, aA4p
AuthorsOpenReview ↗

Further revision following Reviewer aA4p's follow-up

Compared with the first revision, we have further revised the manuscript, following Reviewer aA4p's follow-up suggestions. The additional changes, highlighted in green, are:

  1. Making ROAD’s simple, single-stage, single-objective framework explicit and highlighting its one-step performance against few-step methods.
  2. Restructuring the paper’s narrative to bring out the methodological insights more clearly.
  3. Emphasizing ROAD’s strong diversity retention, also addressing a concern raised by Reviewer YYJG.

We sincerely appreciate this helpful guidance and thank the AC and all reviewers for the time and effort you have devoted to our paper.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.