acceptodds
Withdrawn from ICLR 2027

Bridging Scale-Equivariant Reconstruction and Photorealistic Generation for Blind Face Restoration

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

In blind face restoration, a central challenge lies in reconciling two distinct objectives: scale-equivariant reconstruction and photorealistic generation. Reconstruction models naturally handle arbitrary degradation scales and preserve facial structure, but fail to synthesize realistic high-frequency details. Diffusion models provide powerful generative priors for detail synthesis, yet their fixed-resolution latent sampling process is not inherently compatible with continuous scale variations and often depends on expensive multi-step denoising. To unify the structural reliability of reconstruction with the perceptual realism of generation, we introduce FaceDNO, a novel paradigm that decouples the restoration process into two native mathematical spaces. Specifically, we formulate reconstruction in a continuous function space by employing a resolution-equivariant neural operator to project degraded inputs into a scale-robust structural anchor. We then formulate generation in the discrete latent space, integrating this continuous anchor into a diffusion prior via a manifold-preserving Latent Spatial Feature Transform (Latent SFT). To seamlessly bridge these two spaces, we formulate the operator's reconstruction in the latent space as an isomorphic linear structure to the diffusion forward process. By matching their marginal distributions via Signal-to-Noise Ratios (SNR), we analytically identify the optimal mid-timestep for generative refinement. Experiments demonstrate that FaceDNO effectively combines scale-equivariant reconstruction with photorealistic generation, achieving state-of-the-art fidelity and realism across diverse degradations with one-step diffusion refinement.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Blind face restoration, One-step diffusion, Neural operator, Latent space alignment

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_4wpbOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The manuscript introduces FaceDNO, a paradigm designed to unify structural reconstruction and perceptual generation for face restoration. The framework formulate the resolution-equivariant neural operator's reconstruction as a linear structure isomorphic to the diffusion forward process. By matching marginal distributions via signal-to-noise ratios (SNR), the method analytically determines the optimal intermediate timestep for generative refinement. Experiments reports state-of-the-art fidelity and realism across diverse degradations.

Critical strengths

Unifying neural operators with diffusion priors to bridge structural reliability and perceptual realism is an interesting design. The use of SNR matching to analytically determine the optimal intermediate timestep for generative refinement provides a principled connection between reconstruction and diffusion processes.

Critical weaknesses
  1. I find the overall presentation difficult to follow, including obvious issues like:
  • Inline citations are poorly integrated and lack clear distinction from the main text, severely hurting readability.
  • Ambiguous naming like the term "Mechanism-isolating evaluation", raised as one of the main contributions, appears to be normal ablation studies.
  • Model variants are inconsistently named and insufficiently defined. It is unclear what specific architectural or methodological differences distinguish FaceDNO, FaceDNO-Minimal, FaceDNO-Full, and FaceDNO-Fast (Tables 1, 2, 3, and 4), or Latent SFT versus Latent SFT (full) (Table 8).
  • Several symbols like LDv3D,Lid,LlqconsL_{Dv3D}, L_{id}, L_{lqcons} in Equation 15 are introduced without explanation or definition in the main text.
  • Font sizes in Tables 6, 7, and 8 vary inconsistently, undermining the professional quality of the manuscript.
  1. In addition, FaceDNO necessitates a clearer clarification of its true contribution. In current organization, the manuscript risks appearing as an incremental assembly of existing work. Specifically, the first contribution "Dual-role conditional latent construction", proposes FRONet largely based on NeurOp-Diff without sufficient analysis on its role beyond metrics performance. Regarding the second contributions "First-order SNR calibration", utilizing intermediate diffusion timesteps for one-step denoising, aligning low-quality latents with noisy diffusion latents via LRR, and using SNR for timestep determination are heavily similar to OMGSR, cited several times in the main text. Since the second training stage described in Line 267-269 freezes the FRONet, the overall framework appears largely as utilizing OMGSR as a post-processing for a NeurOp-Diff-like FRONet.
Questions for authors

Given that OMGSR already leverages intermediate diffusion timesteps, LRR for latent alignment, and SNR-based timestep determination, how FaceDNO fundamentally advances beyond an incremental combination of OMGSR and NeurOp-Diff?

Reviewer_aX1oOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper combines multiple strategies to address the problem of blind face restoration, achieving competitive quantitative results across several benchmarks, while also demonstrating clear advantages in terms of qualitative results.

Critical strengths

The qualitative and quantitative results are competitive compared to the baseline.

Critical weaknesses

The issues and concepts described by the author are difficult to grasp. For instance, terms like "conditioning problem" and "the anchor" appear to be new concepts introduced by the author in the context of the blind face restoration problem, yet they are not clearly explained.

It is unclear what the primary contribution of the paper is. Although three contributions are listed, the main one is not explicitly identified. The model is a combination of five different loss functions, it is difficult to determine whether the innovation and core contribution lie in a specific aspect, or if it is merely an aggregation of multiple loss functions.

Questions for authors

The concepts and descriptions in the manuscript are not easily understood, and the primary contributions remain unclear. It appears to be little more than an accumulation of five different loss functions.

Reviewer_EPq9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper does one-step blind face restoration. FaceDNO is a two-stage pipeline. Stage 1 builds a scale-equivariant anchor with a neural operator (FRONet = SRNO querying + Galerkin attention + a high-frequency gated branch). Stage 2 encodes the anchor and the upsampled LQ into an SD2.1 VAE latent, fuses them in latent space, and runs one LoRA UNet step. The optimal timestep is picked by matching the fused latent's SNR to the schedule. They report SOTA on CelebA-Test and three real-world sets with one step, plus lots of ablations.

Critical strengths
  • The results are strong. FaceDNO beats the baseline OSDFace on perceptual metrics. Latency stays competitive (0.82 s), and the Fast variant is even faster than OSDFace.
  • The evaluation is thorough. Table 3 (OSDFace, OSDFace+FRONet, FaceDNO-Minimal, FaceDNO-Full) shows the two-stage conditioning, not the anchor backbone, drives most of the gain.
Critical weaknesses
  • The method rests almost entirely on an integration of different methods. FRONet = SRNO querying + NeurOp-Diff Galerkin attention. The one-step mid-timestep + LRR setup is taken straight from OMGSR. The one genuinely new piece, first-order SNR calibration, is just a dataset-level scalar rescale. The SNR calibration is not critical based on the results in Table 7
  • The scale-equivariant claim is not directly tested, which is the motivation of FRONet. But eval is a standard fixed-512 BFR benchmark. There is no experiment that isolates cross-scale / arbitrary-resolution generalization, where a scale-equivariant operator should win for the right reason.
Questions for authors

Can you give a direct arbitrary-scale / cross-resolution eval (e.g. restoring at resolutions unseen in training) that isolates the benefit of scale-equivariance, instead of only fixed-512?

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

Related papers

Discussion (0)

Sign in to comment.