acceptodds
Under review as a conference paper at ICLR 2027

Perceptual Flow Matching for Few-Step Generative Modeling

Abstract

We propose Perceptual Flow Matching (PFM), a simple yet effective framework for few-step generation in flow-matching models. Rather than performing velocity regression in the conventional VAE latent space, PFM supervises flow matching in a perceptual feature space using pretrained perceptual models. This simple change substantially improves the few-step generation capability of flow-matching models, reducing the number of sampling steps from 35–50 to 4–8 while preserving generation quality. Unlike existing acceleration and distillation approaches, PFM requires neither teacher models nor auxiliary score networks and can be integrated into standard flow-matching training pipelines with minimal modifications. Extensive experiments on image generation, video generation, and image editing tasks demonstrate that PFM consistently produces high-quality results while producing fewer artifacts than existing distillation-based methods. We further show that perceptual supervision shifts the regression minimizer from mean-seeking to mode-seeking, biasing predictions toward on-manifold modes that remain accurate under coarse few-step integration. Our results reveal that standard flowmatching training can naturally yield high-quality few-step generators when supervised in an appropriate representation space. We hope this insight inspires future research into representation-aware objectives for efficient generative modeling.

Primary area:
generative models
Keywords:
diffusion, few-step distillation, aigc

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_7xugOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes Perceptual Flow Matching (PFM). Instead of supervising the velocity prediction using the conventional Euclidean objective in VAE latent space, PFM first converts the predicted velocity into a clean-sample estimate x^0 decodes it using the pretrained VAE decoder, and minimizes a distance between pretrained perceptual features of the predicted and ground-truth decoded images. Experiments are conducted on SD3-Medium for text-to-image generation, Qwen-Image-Edit for image editing, and Wan2.1-1.3B for text-to-video generation.

Critical strengths
  1. The proposed method is conceptually simple and requires only a relatively minor modification to the standard flow-matching pipeline.

  2. The method shows encouraging few-step generation results across text-to-image generation, image editing, and text-to-video generation, suggesting that the empirical benefit is not restricted to a single model or task.

Critical weaknesses
  1. Lack of comparisons with recent strong few-step generation baselines. The paper focuses on few-step generative modeling, but the main comparison is limited to LCM and DMD2. These methods represent relatively earlier generations of few-step/acceleration approaches, while several more recent methods have substantially advanced this area. In fact, the paper itself discusses more recent approaches such as MeanFlow in the related work, but does not include them in the quantitative comparison.

  2. I have concerns about the supervision-space ablation in Figure 3. The performance gap between PFM and all alternative objectives (velocity, x0 pixel, and Huber regression) is surprisingly large, yet the experiment does not appear to fully isolate the supervision space as the only variable. It is unclear whether optimization-related factors such as loss normalization/scaling and other objective-specific hyperparameters were separately tuned for the competing objectives.

  3. The quantitative ablation of the proposed perceptual objective remains limited. Although the paper provides several ablations on supervision spaces and perceptual backbones, these experiments mainly demonstrate that the proposed loss works empirically, rather than providing a thorough quantitative understanding of why it works.

Questions for authors

Please see the Critical Weaknesses.

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

The authors propose PFM, where they replace latent space velocity prediction with perceptual supervision in the decoded RGB estimates. PFM can generate with few-step by finetuning pretrained flow matching models without using a teacher model or other score networks. CFG baking during training can be used optionally.

Critical strengths
  • Simple and reasonable idea. Completely replacing the latent space velocity loss is interesting.

  • Ablations are relevant. Comparing against the velocity or latent x0 prediction show that perception loss in the decoded RGB space is indeed effective.

  • The applicability seems nice. Experiments are conducted across image generation, editing, and video generation using this simple idea seems practical.

Critical weaknesses
  • Limited novelty and validation of CFG baking. Other than the perceptual loss, the CFG baking follows GFT. The actual benefit of GFT+perceptual loss is not fully experimented.
Questions for authors
  • Is the LCM, DMD2 and PFM fully converged at 1000 iterations?

  • Why does FID worsens from 4->8step in Table2?

  • How slow is the training compared with the baselines?

  • Perceptual loss is not a novel loss. What makes this work different from all the other image generation works that uses perceptual loss?

Reviewer_YoZvOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes Perceptual Flow Matching (PFM), a simple change to flow matching in which the authors propose changing the optimization objective from the standard MSE loss in Pixel / VAE latent spaces into a perceptual similarity loss in a pretrained model’s feature space, similar to LPIPS. The authors argue that these losses have an added penalization over blurry off-manifold regions, which pushes the output of a 1-step denoiser closer to the true data manifold, resulting in better few-step generation results. PFM is evaluated on text-to-image, image editing, and video generation settings and shows promising results.

Critical strengths

The paper is clearly written, the motivation is well explained, and the method is simple and easy to implement. The available experimental results, although not comprehensive, are also promising.

Critical weaknesses

The idea of using perceptual losses to improve diffusion and flow-matching models is not new, and multiple prior works [1, 2, 3] have shown that they can greatly improve image quality. These papers seem extremely relevant to the submission, yet there is no discussion of how PFM fits in with these prior works, nor any comparisons against them as baselines.

Furthermore, the experimental evaluation is a bit limited. The authors show results on three different settings: text-to-image generation, image editing, and video generation. However, they only compare against two prior few-step methods, and only in the text-to-image setting, using a single benchmark (COCO 2014 val). Notably, the authors train all methods using an in-house dataset, with little explanation of how it was obtained or created, making reproducibility harder. This also makes it difficult to rerun additional baselines and spend enough effort tuning each method. As a result, it is hard to judge how PFM compares against the numerous few-step methods in the literature. This issue could have been avoided by choosing a more common experimental setting used in prior works, such as those in Tables 1 (ImageNet) and 2 (COCO 2014) of DMD2.


[1] Lin, Shanchuan, and Xiao Yang. "Diffusion model with perceptual loss." arXiv preprint arXiv:2401.00110 (2023).

[2] Ma, Zehong, Ruihan Xu, and Shiliang Zhang. "PixelGen: Improving pixel diffusion with perceptual supervision." arXiv preprint arXiv:2602.02493 (2026).

[3] Berrada, Tariq, et al. "Boosting latent diffusion with perceptual objectives." International Conference on Learning Representations. Vol. 2025. 2025.

Questions for authors
  • Can the authors explain how the in-house dataset was gathered/generated?
  • Could the authors explain why they chose not to use the same experimental settings as DMD2 and instead train on an in-house dataset?
  • Looking at Table 2 of DMD2, where the authors distill SDXL and evaluate it on COCO 2014, both the FID and CLIP scores seem significantly better than those reported in Table 3 (FID: ~20 vs. ~30, CLIP: ~33 vs. ~31). Given that SD3.5-M is a newer and generally more capable model, do the authors have any explanation for these differences?
  • The results in Table 6 seem to suggest that prediction-side CFG baking is the superior choice compared to target-side CFG baking across all metrics. This also makes intuitive sense, since the target-side approach can push the training targets off-manifold, hurting performance. Is this the correct conclusion to draw from these results, or are there specific scenarios where target-side CFG baking might be preferable?
  • Could the authors explain why they chose to define RϕR_\phi the way they did, rather than using the more straightforward definition Rϕ(a,b)=(dϕ(a,m)+dϕ(m,b))/dϕ(a,b)R_\phi(a,b) = (d_\phi(a,m) + d_\phi(m,b)) / d_\phi(a,b)?
  • The discussion in lines 289-302 is a bit confusing. For each anchor image aa, one of its 8 nearest neighbours is chosen as bb. Then, mm is constructed by averaging the 100 nearest neighbours of aa, resulting in a blurry, off-manifold sample. These three images are then used to compute Rϕ(a,b)R_\phi(a,b). However, the earlier definition of RϕR_\phi assumes that m=(a+b)/2m = (a+b)/2, which is confusing. More importantly, looking at the numbers in Table 1, it seems that for perceptual networks such as ConvNeXT, Rϕ≈2R_\phi \approx 2 would imply that dϕ(a,m)≈dϕ(a,b)d_\phi(a,m) \approx d_\phi(a,b). In other words, the distance from aa to a nearby on-manifold sample bb is roughly the same as its distance to the off-manifold sample mm. This seems to suggest the exact opposite of what the authors claim in the text, which is that perceptual losses penalize off-manifold samples more heavily. This needs further clarification.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.