acceptodds
Under review as a conference paper at ICLR 2027

DEYES: Denoise Everything You Ever Saw — Pretrained Visual Priors for Spatiotemporal Monte Carlo Denoising

Abstract

Physically based rendering produces much of the synthetic data that vision models train on, and Monte Carlo sampling sets its cost. A denoiser must reconstruct unbounded linear HDR radiance corrupted by heavy-tailed noise, a problem unlike photographic restoration. We ask whether visual representations learned from photographs transfer to Monte Carlo denoising. DEYES reads the features of a pretrained DINOv3 ConvNeXt encoder out as a multi-scale kernel pyramid that filters linear radiance with non-negative weights. Fine-tuned only on the open Noisebase corpus, it beats OIDN 2.5 and OptiX 9 on PSNR at every sample count from 1 to 512 on 39 scenes from an unseen renderer (Mitsuba 3), and beats OIDN, OptiX and NPPD on every temporal metric on Noisebase video. With its trunk frozen, DEYES beats every baseline on mean PSNR from one eighth of the training data and 9 A40-hours. Controls attribute the gain to the prior: the identical network trained from scratch peaks 1.04 dB lower and never reaches the best production denoiser on mean PSNR; a ConvNeXt pretrained by masked reconstruction on ImageNet-1k recovers most of DINOv3's gain, while DINOv3's own weights with their structure shuffled recover little. Measured end to end, DEYES reaches the best baseline's PSNR with 27–38% less render-plus-denoise time across quality targets from 32 to 512 samples per pixel. Two variants run at about 3 and 10 frames per second at 1920×1024 on one A40. Code, trained models and the Mitsuba evaluation set will be released.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Monte Carlo denoising, transfer learning, pretrained visual representations, self-supervised learning, DINOv3, physically based rendering, HDR image reconstruction, kernel prediction, video denoising, scaling laws

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_sMigOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper propose a new pipeline named DEYES, which adapts a pretrained visual encoder to spatiotemporal Monte Carlo denoising by predicting multiscale filters for linear HDR radiance. The paper evaluates whether image pretraining improves denoising quality, data efficiency, and transfer across renderers.

Critical strengths
  1. The comparisons cover production denoisers and the closest learned baseline across sample counts, with strong gains on both Noisebase and a held-out renderer.
  2. Matched-architecture scratch controls, multiple pretrained initializations, training-budget sweeps, and repeated runs provide unusually useful evidence for the contribution of pretraining.
Critical weaknesses
  1. Cross-renderer temporal generalization. The held-out Mitsuba evaluation consists of static images with history reset, whereas temporal metrics are measured on Falcor sequences; the method's temporal behavior on an unseen renderer remains untested.
  2. Restricted NPPD input protocol. On Mitsuba, NPPD receives repeated pixel-mean buffers instead of its expected per-sample inputs, making its cross-renderer result difficult to interpret.
  3. Approximate high-spp training data. Training examples above 32 spp are synthesized by rescaling 32-spp residuals, which does not reproduce the full distribution of independently sampled high-spp renders.
Questions for authors

Please see the weakness.

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

This paper presents DEYES, a Monte Carlo rendering denoising method that incorporates a pretrained DINOv3 ConvNeXt encoder into a multi-scale kernel-prediction framework. The method aims to transfer pretrained visual representations to rendering denoising to improve image quality, temporal consistency, and computational efficiency. The authors evaluate DEYES on the Noisebase dataset and 39 unseen scenes rendered with Mitsuba 3, comparing it with existing denoising methods. They also conduct controlled experiments to investigate the contributions of pretrained representations, evaluate robustness to missing auxiliary buffers, and analyze the trade-off between denoising quality and rendering time. The results demonstrate promising performance across multiple evaluation settings, suggesting that pretrained visual representations can be beneficial for Monte Carlo rendering denoising.

Critical strengths

The paper presents a systematic study of transferring pretrained visual representations to Monte Carlo rendering denoising. The controlled comparisons between pretrained and randomly initialized models, together with experiments using alternative pretraining sources and shuffled pretrained weights, provide useful evidence for the effectiveness of pretrained visual priors.

Critical weaknesses

The methodological novelty could be clarified further. Since DEYES builds upon an existing multi-scale kernel-prediction framework, the paper would benefit from a clearer explanation of its contributions beyond incorporating a pretrained encoder.

Questions for authors

How consistent are the reported end-to-end efficiency gains across different scenes? In particular, do the conclusions remain similar when unusually expensive scenes are excluded from the analysis?

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

This paper proposes DEYES, a Monte Carlo denoiser that uses a pretrained DINOv3 ConvNeXt encoder, trained on natural photographs, as the feature extractor of a kernel-predicting denoiser.

Trained only on Noisebase, the model outperforms other learning-based denoisers, such as OIDN, OptiX and NPPD, on image and video metrics, both on Noisebase and on Mitsuba scenes.

The method is data- and compute-efficient.

The authors provide comprehensive ablations to attribute the effectiveness to the pretrained visual encoder.

Critical strengths

1, Extensive experiments and well-designed controls for the main claim. The authors compare the encoder trained from scratch, other pretrained initializations (including a ConvNeXt pretrained by masked reconstruction on ImageNet-1k, which shows the gain is not specific to DINOv3), and a DINOv3 trunk with shuffled weights. The shuffled trunk only recovers at most 30% of the gain, so the improvement comes from the learned structure of the pretrained weights. These controls ablate out other factors such as the architecture, the training data and length, the training recipe, the scale of the weights, and the specific pretraining of DINOv3.

Critical weaknesses

1, The writing is often hard to follow. The abstract is dense with numbers and terminology, so the main takeaway is lost. In the introduction, lines 090-092 do not clearly connect to the stated question of whether representations learned from photographs transfer to noisy HDR reconstruction, and lines 095-098 go into method details that belong in Section 3. Several non-standard terms are used without definition (e.g., "the encoder octave", "auxiliary tower" and "sample-count plane"), the operator ⊙\odot is not defined, and the operators SS in Eq. 2 and U(i)U^{(i)} in the coarse-to-fine recombination are described only in words. This does not undermine the results, but it makes the method harder to assess and reproduce.

2, The method section reads like an implementation log: it lists design choices and hyperparameters without explaining the motivation behind them. As a result, it is hard to tell which components are essential to the method and which are incidental engineering choices.

Questions for authors
  1. Mitsuba differs from Falcor/Noisebase in both renderer and scene content, so the Mitsuba results cannot separate renderer shift from scene shift. Both test sets appear to contain scenes named “kitchen” and “bedroom”. Are these the same scene assets? If so, could the authors report results on scenes shared by both renderers, to isolate the effect of the renderer?

  2. The failure cases in Appendix N suggest that performance depends on the type of illumination. Could the authors report per-scene results grouped by lighting conditions or lighting difficulty, to show whether DEYES wins across scene types rather than only on the 39-scene average?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.