acceptodds
Under review as a conference paper at ICLR 2027

FREPix: Frequency-Heterogeneous Flow Matching for Pixel-Space Image Generation

10 views

Abstract

Pixel-space diffusion has re-emerged as a promising alternative to latent-space generation because it avoids the representation bottleneck introduced by VAEs. Yet most existing methods still treat image generation as a frequency-homogeneous process, overlooking the distinct roles and learning dynamics of low- and high-frequency components. To address this, we propose , a quency-heterogeneous flow matching framework for el-space image generation. explicitly decomposes generation into low- and high-frequency components, assigns them separate transport paths, predicts them with a factorized network, and trains them with a frequency-aware objective. In this way, coarse-to-fine generation becomes an explicit design principle rather than an implicit behavior. On ImageNet class-to-image generation, achieves competitive results among pixel-space generation models, reaching 1.91 FID at and 2.38 FID at , with particularly strong behavior in the low-NFE regime.

Primary area:
generative models
Keywords:
pixel-space generation, diffusion model, flow matching, frequency heterogeneity

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_sJcCOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces FREPix, a frequency-heterogeneous flow-matching framework for pixel-space image generation. FREPix uses an orthonormal Haar wavelet transform to decompose images into low- and high-frequency states, assigns them distinct interpolation schedules, and explicitly factorizes prediction between a low-frequency DiT structure predictor and a lightweight high-frequency decoder conditioned on the predicted structure. It further applies time-dependent frequency weighting to the flow-matching objective while preserving the same population-optimal velocity field, supplemented by REPA and LPIPS losses. On class-conditional ImageNet generation, FREPix achieves competitive results at both 2562256^2 and 5122512^2 resolutions, including FIDs of 1.91 and 2.38, respectively. Ablations examine the interpolation schedules, frequency weighting, and auxiliary losses, while additional experiments show particularly strong performance under limited training budgets and low-NFE sampling.

Critical strengths
  1. A coherent frequency-heterogeneous formulation. FREPix consistently incorporates frequency heterogeneity into the transport path, prediction targets, architecture, and training objective. This is more consequential than merely adding frequency-aware features or an auxiliary loss: it turns coarse-to-fine generation into an explicit modeling principle while retaining an exact, lossless pixel-space representation through an orthonormal DWT.

  2. Strong performance in practically important efficiency regimes. FREPix is particularly competitive under limited training budgets and low-NFE sampling, while requiring fewer GFLOPs than most comparable pixel-space models. These results suggest that the proposed inductive bias improves not only final generation quality but also optimization and sampling efficiency, which substantially strengthens the practical significance of the work.

Critical weaknesses
  1. The contribution of the proposed frequency-heterogeneous design is not sufficiently isolated. The reported performance is heavily influenced by REPA and LPIPS—for example, the ablation FID improves from 53.78 to 13.85 when both are added—while the paper lacks a controlled, component-wise comparison of joint versus explicitly decoupled prediction, homogeneous versus heterogeneous paths, and uniform versus frequency-aware weighting. Consequently, it is difficult to determine how much of the gain is attributable to FREPix itself rather than auxiliary supervision.

  2. Important inconsistencies undermine reproducibility and confidence in the empirical claims. The paper alternately reports Euler and Heun as the solver for the main FREPix-XL results, gives inconsistent CFG intervals, and the sampling pseudocode appears to update from xpredx_{\mathrm{pred}} rather than the current state xtx_t, contrary to a standard Euler step. Since sampling configuration directly affects FID and low-NFE performance—one of the paper’s central claims—these ambiguities are consequential to the overall assessment.

Questions for authors

Clarification requests

  1. What sampling procedure was used for the main FREPix-XL results: Euler or Heun? Please also clarify the CFG interval and the actual number of network evaluations. In Algorithm 2, should the Euler update start from the current state xtx_t, rather than from xpredx_{\mathrm{pred}}?

  2. Were the baselines in Tables 1 and 3 trained or evaluated with the same REPA and LPIPS supervision, time sampler, CFG configuration, and computational budget as FREPix? If not, please clarify which components and settings differ.

Additional evidence that would materially affect the assessment

  1. Can the authors provide a controlled component-wise ablation comparing: (i) joint versus explicitly frequency-decoupled prediction, (ii) homogeneous versus heterogeneous transport paths, and (iii) uniform versus frequency-aware loss weighting, while keeping the architecture size and auxiliary losses fixed? The current ablations do not isolate the contribution of the paper’s central design choices.

  2. Can the authors report matched comparisons of FREPix, JiT, and DeCo with identical REPA and LPIPS supervision? Given that these auxiliary losses reduce FREPix’s FID from 53.78 to 13.85, such a comparison would materially affect whether the reported gains can be attributed to frequency-heterogeneous modeling.

Reviewer_AK6NOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a pixel-space flow matching framework that accounts for the difference of low and high frequence components. Specifically, they do so by using separate transport path, low-freq and high-freq network, and frequency-aware objective. This allows them to learn coarse-to-fine generation more explicitly. Empirically show that they achieve competitive FID scores on 256 and 512 resolution.

Critical strengths
  • The motivation of frequency discrepancy is validly sound and justifiable but it is not fundamentally new.
Critical weaknesses
  • The concept of learning a model on separate frequency roles like low and high are very common and standard. Same for the observation of heterogeneity (or simply asymmetric dynamics) between low and high frequency. Though I acknowledge the technical components the method introduce to better exploit this, the method is only considered as incremental work.
  • Ablation of using shared schedule for both frequencies is needed. This mean that both frequencies are jointly leant through a single backbone like in WDM (Phung et al) paper, very simple version, no low-freq conditioning and separate network.
  • Table 6. the performance of the base method is super worse with only the frequence loss, this yields a question whether the advantage is coming from LPIPS and REPA or this method is only working if it paired with other strong losses.
  • Table 3: there is no much gap between DeCo and the method. The only thing that the method win over DeCo and related work is the few-step quality.

Exposition:

  • Related work: it is not clear what is different between this work and Deco since they both look at frequency perspective and are pretty much similar. Pls justify. I saw fig. 4 illustrate the paradigm comparison, it should be mentioned in the related work. and
  • Figure 7 seems to me a supplementary material.
  • The ablation section needs work. Why include the equations and their details in ablation. This section only to show the results, that's it!
Questions for authors

See the weakness.

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

The authors observe that high-frequency and low-frequency components exhibit drastically different properties in diffusion models for natural images. Accordingly, they propose to explicitly partition the image domain into two separate frequency bands and apply different schedules for diffusion generation respectively. They claim that this approach yields improved training efficiency and generation quality, with validations conducted on 256×256 and 512×512 image generation tasks.

Critical strengths

The phenomenon observed by the authors is well known, and the solution approach is simple and intuitive. The authors provide some theories to support their model. The experimental results to some extent validate their claim.

Critical weaknesses

Main weakness:

  1. The improvement in training efficiency brought by the REPA/LPIPS loss is questionable. As proposed in the previous paper PixelGen, the REPA+LPIPS loss itself can significantly improve training efficiency, which makes the reported training efficiency gain in this work unconvincing. I notice that the authors have conducted ablation experiments on REPA and LPIPS regarding performance, but ablations targeting the effect of REPA and LPIPS on training efficiency are still required.

  2. latent space vs. image space vs. freq. space. The authors essentially adopt a deco-like scheme in the freq. space for diffusion. Apart from training efficiency, are there any specific advantages of using the freq. space? Especially considering that the latent space from prior work can be regarded as an upper bound of the freq. space?

  3. Lacking experiments on Text-to-image experiments significantly reduces the values to the research community since class-to-image generation is nearly addressed well.

Minor weakness:

  1. In Equation 2 and Fig. 3, when constructing LF and HF, the authors do not use two independent noise signals; instead, they split a single complete noise into two parts via DWT. I wonder whether these two split components can still be regarded as Flow Matching trajectories. How large is the performance gap compared with training using two independent noise signals?

  2. I personally consider several theorems in this paper redundant, and the theory seems to be added merely to increase complexity. The authors should clarify the necessity of these theorems and explain what guiding effects they have on the final experimental performance or experimental design.

  3. Comparisons for dual-trajectory diffusion are required. Latent forcing is a diffusion model that implements dual trajectories in both the latent and image domains, and the authors should compare against it.

Questions for authors

See above. Q1/Q2/Q3 is the main concerns.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.