- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper proposes Flicker Diffusion, which replaces standard spatially independent Gaussian noise with spatially correlated Gaussian noise whose covariance decays as a power law with the Manhattan distance between pixels. The method uses the same covariance consistently in both the diffusion process and a corresponding whitened training objective. Its motivation is to better match the spatial and spectral correlations of natural images and thereby make the denoising dynamics across different frequency modes more synchronized. Empirically, Flicker Diffusion can achieve, in the low-NFE (e.g. 20) regime, sample quality comparable to white-noise diffusion using substantially (e.g. 5x) more network evaluations. This advantage appears to saturate as the sampling budget increases.
Critical strengths
-
Simple and interesting design.
The paper introduces a simple way to incorporate structured spatial correlations into diffusion training while keeping the model architecture unchanged. The same formulation can be used consistently during training and sampling, without requiring an additional distillation or post-training stage. -
Careful empirical validation.
The experiments are relatively thorough, with multiple datasets, paired training seeds, matched training budgets, and controlled sampler selection. The paper also reports unfavorable settings, including cases where Flicker does not outperform the white-noise baseline, which makes the empirical conclusions more credible.
Critical weaknesses
-
The practical advantage appears limited to the low-NFE regime.
Flicker performs well with few sampling steps, but its performance saturates as NFE increases, while the white-noise baseline continues to improve and can outperform it at higher NFE. This limits the performance ceiling of the method, which is important because a strong high-NFE model can later serve as a better teacher for few-step distillation. -
The validation is limited to small pixel-space diffusion models.
The experiments are mainly conducted on CIFAR-10 and 64×64 datasets with U-Nets. It remains unclear whether the same covariance design transfers to modern latent diffusion or DiT-based models, where the representation is higher-dimensional and has very different spatial and spectral statistics.
Questions for authors
- The main models are trained for 90k updates. Have the authors tried training both the white-noise and Flicker models for longer? It would be useful to know whether the observed performance ceiling remains with a larger training budget.
- Could the authors improve the presentation to make the experimental settings easier to follow? For example, T=150/NFE40 is a bit of unnatural.
- I acknowledge the effectiveness of the idea in short horizon, but the current evaluation setting feels somewhat dated, with most experiments conducted on small-scale pixel-space U-Nets and mainly evaluated by FID. Could the authors provide some results on a more modern setting, such as latent diffusion or DiT-based models at higher resolution, to show whether the method still transfers? (If computationally not applicable, I'd like to hear the discussion or attempts on it.)
- The paper evaluates DDIM and DPM-Solver++, but several stronger modern solvers are now available. Could the authors also compare Flicker with solvers such as UniPC and iPNDM [1] under the same NFE budget to see whether the low-NFE advantage still holds?
[1] UniPC: A Unified Predictor-Corrector Framework for Fast Sampling of Diffusion Models