acceptodds
Under review as a conference paper at ICLR 2027

FlickerDiff: Uncertainty-Controlled One-Step Diffusion for Flicker Removal

Abstract

Spatially non-uniform degradation makes it difficult for image restoration models to balance fidelity and generative recovery. Single-image flicker removal is a representative example of this problem. Flicker bands can severely attenuate text, edges, and object structures in some regions while leaving nearby content largely intact. Existing diffusion-based restoration methods often apply a uniform noise level or restoration strength across the image. They may therefore under-recover severely degraded regions or unnecessarily modify reliable content. We propose FlickerDiff, a one-step pixel-space diffusion model that adjusts this balance according to local reliability. A lightweight auxiliary network predicts an uncertainty map, which determines the spatially varying transport variance in the flow-matching endpoint distribution. Reliable regions follow narrow paths anchored to the observation, while severely degraded regions receive a larger stochastic transport budget and can rely more strongly on the pretrained generative prior. The same uncertainty map is injected into the generator through token-aligned FiLM modulation. We adapt a pretrained 8B-parameter pixel-space Diffusion Transformer in its velocity-prediction space to build the one-step generator. Adversarial detail supervision and a lightweight post-fusion refiner further improve perceptual quality and output fidelity. Experiments show that FlickerDiff achieves the strongest reconstruction fidelity and perceptual quality among the compared methods. At inference, the full pipeline consists of uncertainty estimation, a single Pixel-DiT forward pass, and lightweight post-fusion refinement, without iterative diffusion sampling.

Primary area:
generative models
Keywords:
Image Restoration, Diffusion Models, Generative Restoration, Uncertainty Estimation, Pixel-level Diffusion Models, One-step Generation, Low-level Vision

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

The proposed FlickerDiff first estimates an uncertainty map from the flicker-corrupted image, which is then used to control the noise distribution in flow-matching and guide the FiLM modulation for restoration.

Critical strengths

The experimental results show that FlickerDiff outperforms existing methods on the BurstFlicker-S dataset across all metrics. Ablation studies further demonstrate the effectiveness of each component.

Critical weaknesses

(1) The contribution appears to be mainly a combination of existing components. FlickerDiff integrates uncertainty estimation, spatial noise scaling, FiLM modulation, adversarial training, and a NAFNet-style refiner on top of an 8B pretrained model. Meanwhile, post-fusion contributes most of the PSNR improvement in the component ablation study. In Table 15, adding the uncertainty map and FiLM improves PSNR by only 0.641 dB, whereas post-fusion achieves a much larger gain of 1.786 dB. This indicates that the final reconstruction performance may rely heavily on the additional refinement module rather than the proposed uncertainty-guided generation process.

(2) Most compared diffusion-based methods, such as PiSA-SR, OSEDiff, ResShift, and SinSR, were originally designed and trained for SR tasks. However, the paper does not clearly describe how these methods are adapted to flicker degradation, including the training data, loss functions, and fine-tuning strategies. Without these details, it is difficult to determine whether the comparisons are conducted under a fair setting. Moreover, the pretrained HiDream-O1 backbone contains 8B parameters, which is substantially larger than the backbones used in these competing methods.

Questions for authors

The proposed method uses an uncertainty map to modulate the noise distribution and compares it with fixed noise. However, is noise injection necessary in this single-step flow-matching setting? Existing single-step diffusion-based methods, such as OSEDiff, directly take the low-quality input without adding noise. The inferior performance of fixed noise may result from inappropriate noise magnitude or the limited number of inference steps, rather than from the fixed-noise strategy itself. Therefore, it would be valuable to clarify the motivation for introducing noise in a single-step flow-matching framework and whether uncertainty-guided noise modulation provides benefits beyond simply reducing excessive noise in relatively certain regions.

Reviewer_gU7JOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

In this paper, the authors propose FlickerDiff, which is an uncertainty-guided one-step diffusion model for the removal of flicker in a single image. A Restormer-based uncertainty map is used by them to control the spatial noise in a pretrained Pixel-DiT. The restoration quality is further improved by FiLM conditioning and by post-fusion refinement. The experiments on BurstFlicker-S show strong results, both in reconstruction fidelity and in perceptual quality.

Critical strengths

It is an interesting problem that the paper addresses, namely to balance content preservation and restoration in the regions with different levels of degradation.

The proposed uncertainty-guided endpoint modulation is reasonable and constitutes a useful extension to one-step diffusion restoration.

Strong quantitative results are achieved by the method, and it outperforms the compared methods across several image quality metrics.

Detailed ablation studies and hidden-feature analyses are provided in the paper to support the proposed design.

Critical weaknesses

The technical novelty is somewhat limited, because the uncertainty-guided perturbation has already been explored in previous diffusion-based restoration methods.

The final performance gain relies heavily on the post-fusion. As shown in Table 15, it is a much larger PSNR improvement that post-fusion provides than uncertainty guidance.

The improvement over uniform noise is relatively small (0.127 dB in Table 2). More experiments are necessary to confirm the benefit of the spatial uncertainty control.

The evaluation is limited to one main dataset. Additional tests on different flicker conditions would better demonstrate the generalization.

Questions for authors

N/A

Reviewer_k3hdOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper tackles single-image flicker removal on BurstFlicker-S by using an uncertainty map to spatially modulate the starting-state noise in a one-step pixel-space DiT. The map, derived from a frozen Restormer output-input difference, scales per-pixel noise and is also injected via FiLM conditioning. A refiner is applied after the coarse generator. The deployed system achieves 29.00 dB PSNR and 0.067 LPIPS, ahead of eight baselines, with ablations comparing real, uniform-energy, random, and inverted maps.

Critical strengths
  1. Clear problem framing with direct task relevance. Using a spatial reliability signal to balance local completion and fidelity preservation is well motivated for severe flicker.
  2. Strong system-level results. The final model achieves 28.998 dB PSNR and 0.067 LPIPS, ahead of all baselines. The coarse generator alone already reaches 27.290 dB, demonstrating competitive generation quality before refinement.
  3. Substantial verification effort. Energy-matched uniform controls and region-stratified evaluations are appropriate. Hidden-state analyses including token replacement interventions provide evidence beyond final metrics.
Critical weaknesses

W1. The core claim is not isolated: whether the map should control noise or act as a condition.

The "uniform noise + FiLM4" control is missing. Of the 0.68 dB endpoint gain over fixed noise, 0.55 dB comes from energy-matched uniform noise; spatial alignment adds only 0.127 dB. In high-uncertainty regions, real and inverted maps score nearly identically (25.930 vs 25.924). Compared to the inverted map, the real map's advantage appears primarily in low-uncertainty regions.

W2. Limited novelty relative to UPSR, and the map's contribution is not separated from the refiner.

UPSR already uses an auxiliary network's residual as uncertainty and scales per-pixel starting-state noise accordingly. Section 2.4 cites UPSR but does not discuss the mechanistic overlap or provide a direct comparison. Table 15's progressive ablation (Base → +Map → +GAN → +refiner) shows the refiner contributes +1.79 dB, but this does not isolate the map's independent effect, as the coarse generator's Low-U PSNR (26.80) falls below Restormer and ResShift before refinement.

Questions for authors
  1. Does the DiT receive B anywhere in the "real map" variant (Table 2) other than through the noise? Does the best uniform σ₀ close the 0.127 dB alignment gap?
  2. Was "HiDream multi-step" in Table 1 fine-tuned on BurstFlicker-S, and using what data/iterations?
  3. How is region PSNR computed (pixel vs token masks, per-image vs pooled)? Why do both region PSNRs sit below full-image PSNR in Table 7?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.