acceptodds
Withdrawn from ICLR 2027

Curriculum-Driven Degradation-Aware Diffusion Transformers for Real-World Old-Photo Face Restoration

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Real-world old-photo face restoration is severely hindered by the complexity of unknown physical damages and the synthetic-to-real domain gap. While recent Diffusion Transformer (DiT) architectures exhibit powerful generative priors, their static training paradigms treat all degradation levels equally. This inherently leads to a sub-optimal trade-off: over-smoothing on mild degradations and severe identity hallucination on complex physical damages. In this paper, we propose CurrDiT, a curriculum-driven DiT that dynamically adapts to degradation severity. First, we introduce a feedback-driven curriculum sampling strategy built upon a generative degradation process, which dynamically reweights degradation difficulty according to the evolving training state to progressively shift training from simple degradations to complex physical damages, thereby guiding the model to gradually unlock its untapped restoration potential. Second, CurrDiT adopts a dual-stream architecture that integrates degradation-aware local repair with global appearance guidance, where the Damage Restoration Branch (DRB) leverages damage-specific local cues to restore structurally corrupted regions, while the Global Degradation Restoration Branch (GDRB) exploits reliable non-damaged regions to enhance global appearance consistency. We further develop a generative degradation synthesis pipeline tailored to old-photo face restoration and construct a dedicated VintagePortraits benchmark for real-world evaluation of naturally aged old-photo portraits. Extensive experiments on old-photo face and blind face restoration benchmarks demonstrate that CurrDiT achieves superior perceptual quality, structural fidelity, and identity consistency.

Primary area:
generative models
Keywords:
Face Restoration, Diffusion Models, Curriculum Learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_DtcWOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors propose CurrDiT, a curriculum learning-based DiT model capable of dynamically adapting to varying degrees of degradation. Additionally, they developed a generative degradation synthesis pipeline tailored specifically for restoring faces in old photographs and constructed a dedicated benchmark, VintagePortraits, for evaluating naturally aged, authentic portrait photos. Extensive experiments demonstrate that this method offers advantages over existing approaches.

Critical strengths
  1. A generative degradation synthesis pipeline tailored specifically for restoring faces in old photos is valuable.
  2. The performance of this method appears to be superior to that of existing methods.
Critical weaknesses
  1. It is difficult to determine whether the performance improvement stems from the authors' design or from the use of a larger base model, given that this method employs a larger base model (FLUX-1) compared to the baseline methods. Therefore, the authors should compare their approach against the latest methods based on FLUX or on base models larger than FLUX—such as LucidFLUX (based on FLUX-1) and ODTSR (based on Qwen).
  2. As a method requiring supervised training rather than being training-free, it appears to offer no advantages in terms of inference speed or the total/trainable number of model parameters with existing training-based methods.
  3. There are too few methods for identity assessment; the scope should include all the methods listed in Table 1, as well as the newly added methods based on larger base model patterns.
  4. To my knowledge, SSDiff appears to inevitably alter the skin tone of the input face based on pseudo-label references; however, the SSDiff outputs shown in Figure 3 (using the authors' proposed dataset) do not exhibit any skin tone changes. Did the authors perhaps not use pseudo-labels? If that is the case, is the quantitative comparison with SSDiff presented in Table 1—based on the authors' proposed dataset—accurate?
  5. Architectural innovation is limited; it merely incorporates degradation awareness into the dual-stream framework, representing an incremental advancement.
Questions for authors

See Weaknesses

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

This paper proposes CurrDiT, a curriculum-driven diffusion transformer for real-world old-photo face restoration. The method introduces an adaptive curriculum learning strategy that combines prior degradation difficulty with training-loss feedback to progressively emphasize challenging samples. It further develops a dual-branch architecture, where the Damage Restoration Branch focuses on local structural damage and the Global Degradation Restoration Branch exploits undamaged regions to improve global appearance consistency. The authors also design a synthetic degradation pipeline and construct the VintagePortraits benchmark. Experiments on old-photo and blind face restoration datasets demonstrate improvements in perceptual quality and structural fidelity over several existing methods.

Critical strengths
  1. The paper addresses a challenging and practically relevant problem in real-world old-photo face restoration, where complex physical damages and global degradations must be handled simultaneously. The proposed separation of local damage restoration and global appearance guidance is well motivated and provides a reasonable framework for addressing these distinct challenges.
  2. The experimental results are promising across multiple old-photo and blind face restoration benchmarks. CurrDiT achieves competitive or superior performance on several perceptual quality metrics, while the ablation studies demonstrate the effectiveness of the proposed restoration modules. The additional VintagePortraits benchmark also contributes to evaluating restoration performance on naturally aged photographs.
Critical weaknesses
  1. The methodological novelty and the contribution of adaptive curriculum learning are insufficiently established. The proposed curriculum primarily combines existing loss-based difficulty estimation, EMA feedback, and a predefined easy-to-hard schedule, while DRB and GDRB rely on conventional attention mechanisms. More importantly, Table 2 shows that adaptive curriculum learning provides only marginal improvements, whereas most gains come from DRB. Comparisons with stronger curriculum learning and loss-aware sampling baselines are necessary to demonstrate the distinct value of the proposed approach.
  2. The experimental validation does not adequately support the central claims regarding real-world restoration and identity preservation. Both DRB and GDRB depend on damage masks, but their availability and generation during real-world inference remain unclear. Furthermore, identity preservation is evaluated only on synthetic paired data, while real-world evaluations primarily rely on perceptual quality metrics that cannot reliably detect identity hallucination. These limitations raise concerns about the method's practical applicability and its ability to preserve facial identity under complex real-world degradations.
Questions for authors

same to weakness

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

This paper introduces CurrDiT, a diffusion-Transformer framework designed for real-world old-photo face restoration. By employing adaptive curriculum learning and a dual-branch architecture, the framework optimizes both the repair of local structural damage and the recovery of global appearance across varying degrees of degradation. Experimental results demonstrate that this method achieves superior performance in terms of restoration quality and identity preservation.

Critical strengths
  1. Propose a feedback-driven adaptive curriculum learning strategy that dynamically adjusts sampling weights across degradation difficulty levels, thereby improving restoration results in complex degradation scenarios.
  2. Employ DRB and GDRB to address local physical damage and global appearance degradation, respectively, which helps enhance restoration quality and structural consistency.
  3. Construct the VintagePortraits dataset—a collection of real old-photo face images—and demonstrate competitive performance across various benchmarks for old-photo face restoration and blind face restoration.
Critical weaknesses
  1. Both DRB and GDRB rely on damage masks; however, the paper does not specify how these masks are obtained during the inference phase for real-world photographs, which limits the practical utility of the proposed method.
  2. In Table 2, the performance gains attributed to adaptive curriculum learning are relatively limited—insufficient to definitively establish it as a key factor in the overall performance improvement.
  3. The existing ablation studies lack a comprehensive, independent evaluation of the various combinations of DRB and GDRB, making it difficult to accurately assess the individual contributions of each module and the extent to which they complement one another.
  4. The paper primarily employs no-reference perceptual quality metrics, while identity consistency is evaluated solely on synthetic paired data; this is insufficient to substantiate claims regarding identity preservation in real-world scenarios involving severe damage.
  5. The proposed method relies on a powerful pre-trained FLUX.1-Kontext-dev model, yet the paper lacks adequate comparative experiments using the same backbone network, making it challenging to disentangle the performance gains derived from the pre-trained priors versus those contributed by the proposed method itself.
Questions for authors
  1. How exactly are damage masks obtained during inference for real-world photographs? Is it possible to provide a fully automated inference pipeline, and can the impact of false positives and false negatives in mask detection on restoration performance be quantified?
  2. According to Equation (3), when the training progress p=0.5p=0.5, the weights for all samples become 0.5. How do the authors explain the fact that difficulty-aware sampling becomes independent of difficulty scores at this stage? Is this design truly superior to traditional curriculum learning?
  3. GDRB uses an attention mask to mask damaged conditional tokens for mainstream queries, whereas DRB injects damage-aware local key-value tokens into the attention computation. Could the authors clarify whether a potential conflict exists between these two components, and provide an analysis of the underlying mechanism or experimental validation?
  4. For authentic old photographs where original identity information is severely missing, how do the authors demonstrate that the model restores the actual original faces, rather than generating plausible-looking faces that are merely hallucinations of the generative model? Could the authors provide an evaluation of identity consistency or validation using real-world paired data?
  5. Is it possible to compare standard LoRA fine-tuning, prior-only curriculum learning, and the full CurrDiT framework—using the same FLUX backbone, training data, and computational budget—while reporting the additional parameter count, inference latency, and VRAM consumption?

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

Related papers

Discussion (0)

Sign in to comment.