acceptodds
Under review as a conference paper at ICLR 2027

Negative Restoration via the Heterogeneity of Channel-Induced Geometric Features

Abstract

Color film negatives are vulnerable to channel-heterogeneous degredation caused by long-term chemical aging, where the blue (B) channel is often substantially more corrupted than the relatively intact red (R) and green (G) channels. Existing restoration methods typically process degraded RGB images holistically and therefore fail to explicitly exploit such cross-channel structural asymmetry. In this work, we propose a geometry-guided approach for negative-film restoration based on a key observation: different color channels depict the same underlying scene geometry and should yield consistent geometric features extracted from the pre-trained 3D foundation model in reliable regions, whereas deterioration introduces cross-channel feature discrepancies. We exploit these discrepancies to localize unreliable B-channel regions and guide selective restoration using information from the other two intact channels while preserving already reliable content. We further utilize the geometric features to facilitate reference-conditioned color mapping by identifying geometrically and semantically related regions between the restored negative and a reference printed image, enabling region-aware color mapping guided by the reference. Extensive experiments demonstrate that our approach achieves superior negative-film restoration and more coherent color mapping than combinations of existing restoration and color-mapping methods.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Image Restoration, Color Mapping

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_DpeHOpenReview ↗
Rating
2(−1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a two-stage geometry-guided framework for restoring deteriorated color film negatives and mapping them to printed color photographs. The authors observe that while chemical aging disproportionately degrades the blue (B) channel of color negatives, all three RGB channels depict the same underlying scene geometry. In Stage 1 (Restoration), the method replicates each RGB channel into a 3-channel view and passes them jointly through a frozen 3D foundation model, Depth Anything 3 (DA3). Cosine distances between intermediate B-channel features and a learned linear combination of R/G features yield a spatial geometric inconsistency map UU, which modulates residual updates in a NAFNet-based U-Net ("DA3-guided NAFBlock") to restore the B channel. In Stage 2 (Color Mapping), frozen DA3 token features are used to compute a cross-view affinity matrix between the restored negative and a reference printed photo from the same film roll, guiding spatially adaptive color transfer via feature aggregation. Experiments on the BlueNeg dataset (augmented with synthetic B-channel fading) show consistent quantitative and qualitative improvements over general/all-in-one image restoration models and reference-based color transfer baselines.

Critical strengths
  1. Well-Motivated Cross-Channel Geometric Prior: Re-purposing intermediate features from a multi-view 3D foundation model (DA3) by treating individual RGB color channels as aligned "views" is a clever and well-motivated formulation. Unlike direct pixel differences—which are confounded by natural object color variations—intermediate geometric features remain invariant to legitimate channel intensity differences while exposing structural degradation ( convincingly demonstrated in Fig. 3 and the pixel/DINOv2 ablations in Table 1).
  2. Unified Pipeline Design & Strong Empirical Gains: Using the same frozen foundation backbone to solve both intra-image channel discrepancy (Stage 1) and cross-image region correspondence (Stage 2) provides a cohesive architectural narrative. The method achieves notable empirical improvements (+0.70 dB PSNR over R2R in restoration; +1.91 dB PSNR over CAP-VSTNet in color mapping) while preserving sharp local structure.
Critical weaknesses
  1. Unaddressed Failure Mode in Per-Image Min–Max Normalization (NmmN_{\text{mm}}): In Eq. (3), the geometric inconsistency map U(q)U(q) is normalized via per-image min–max scaling onto [0,1][0, 1]. Consequently, even if a negative has near-zero degradation—or uniform global degradation across the entire B channel—UU is forced to span the full [0,1][0, 1] range, artificially amplifying residual updates in relatively clean regions. Furthermore, Eq. (1) and the architectural design strictly assume R and G are intact and only predict B^\hat{B}; in real archival films where G (magenta layer) or R (cyan layer) also suffer localized defects or scratches, the reference feature DRGD_{RG} will invert the guidance and corrupt the restoration.
  2. Missing Comparison to the Native Domain Baseline & Computational Overhead Analysis: While the paper trains and evaluates on the BlueNeg dataset (Liu et al., ICCV 2025) and compares against generic RGB restoration networks (e.g., Restormer, PromptIR, R2R), it omits a direct quantitative comparison against the original baseline method proposed in the BlueNeg paper itself, which was specifically tailored for channel-heterogeneous film degradation. Additionally, running a 39-layer 3D foundation model (DA3) across 3 replicated views in Stage 1 and 2 views in Stage 2 introduces massive memory/FLOP overhead compared to a lightweight width-32 NAFNet, yet no runtime, parameter count, or memory comparisons are reported.
Questions for authors
  1. Comparison with BlueNeg Baseline: Why was the restoration method from Liu et al. (2025) (which introduced the BlueNeg dataset and task) omitted from Table 1? How does your DA3-guided NAFNet compare against their domain-specific method?
  2. Computational Complexity: Could you provide a table comparing inference runtime (ms/image), peak GPU memory, and total/trainable parameters of your pipeline (including the 3-view and 2-view DA3 forward passes) against baselines like R2R and Restormer?
  3. Sensitivity to Clean/Uniformly Degraded Inputs: Because NmmN_{\text{mm}} in Eq. (3) forces max⁡(U)=1\max(U)=1 and min⁡(U)=0\min(U)=0 for every image, how does the network behave when presented with an already-clean negative or a globally faded B-channel where relative spatial contrast in UU is pure noise?
  4. Cross-Scene Reference Discrepancies: In Stage 2, when the reference print from the same film roll depicts a completely different scene with minimal semantic/geometric overlap (e.g., an indoor portrait vs. an outdoor landscape), does the softmax affinity matrix MM (Eq. 5) with fixed temperature τ\tau genuinely become diffuse, or does it still peak on arbitrary false-positive patches? What value of τ\tau is used?
11 comments

Thank you for the positive and insightful feedback. We address the concerns below and refer to the revised PDF, where edits are marked with blue text:

 

Q1 Comparison with BlueNeg Baseline & W2

While the paper trains and evaluates on the BlueNeg dataset (Liu et al., ICCV 2025) and compares against generic RGB restoration networks (e.g., Restormer, PromptIR, R2R), it omits a direct quantitative comparison against the original baseline method proposed in the BlueNeg paper itself, which was specifically tailored for channel-heterogeneous film degradation. Why was the restoration method from Liu et al. (2025) (which introduced the BlueNeg dataset and task) omitted from Table 1? How does your DA3-guided NAFNet compare against their domain-specific method?

Based on our understanding, the BlueNeg paper introduced the negative dataset and its evaluation protocol, but did not propose a new domain-specific restoration architecture. In their evaluation, NAFNet and Restormer were used as supervised baselines by training their official implementations on the BlueNeg training split with the estimated pseudo ground truth.

Restormer is already included in Table 1, retrained on our augmented BlueNeg training set under the same evaluation protocol. For completeness, we also retrain the original NAFNet under identical settings. Its performance is reported below and compared with our DA3-guided NAFNet in following table. Our method outperforms the original NAFNet on seven of the eight metrics, supporting the effectiveness of the proposed cross-channel geometry-guided restoration mechanism beyond a standard restoration backbone. We will include this comparison in Table 1 of the revised paper.

MethodBlue Channel LPIPS↓Blue Channel PSNR↑Blue Channel SSIM↑Blue Channel FID↓Overall LPIPS↓Overall PSNR↑Overall SSIM↑Overall FID↓
NAFNet0.036428.51850.977721.49200.022233.28970.99268.0713
Ours (DA3 Guidance)0.029529.56050.983321.31990.021034.33170.99448.9825

 

Q2 Computational Complexity & W2

Additionally, running a 39-layer 3D foundation model (DA3) across 3 replicated views in Stage 1 and 2 views in Stage 2 introduces massive memory/FLOP overhead compared to a lightweight width-32 NAFNet, yet no runtime, parameter count, or memory comparisons are reported. Could you provide a table comparing inference runtime (ms/image), peak GPU memory, and total/trainable parameters of your pipeline (including the 3-view and 2-view DA3 forward passes) against baselines like R2R and Restormer?

We report computational complexity under a unified inference-only protocol on a single NVIDIA A100-SXM4-80GB GPU. All methods are evaluated with batch size 1. Models are first loaded and moved to GPU; checkpoint loading time is excluded, and we run 3 warm-up forwards on the first image (not recorded). Our model does not require running all 39 layers of DA3. Instead, it only uses the first 13 layers for restoration and the first 19 layers for color mapping, as features extracted from intermediate layers already provide sufficient guidance for both tasks. Reported latency is the per-image mean ± std. For restoration, despite a relatively large total parameter count due to the pretrained DA3 backbone, our method requires the fewest trainable parameters among all compared restoration methods while achieving the third-fastest inference time. We will add this experiment to Appendix D.1.

ModelParams (M)Trainable (M)Inference (s/img)↓Peak GPU (MB)↓
Restormer26.126.10.936 ± 0.1466,918 ± 1,130
MambaIRv222.722.715.155 ± 2.59418,972 ± 3,126
UniRestore1050.2100.60.617 ± 0.1887,308 ± 524
RDBM36.336.35.975 ± 1.0316,270 ± 1,763
AirNet8.97.60.896 ± 0.1513,475 ± 571
PromptIR35.635.61.062 ± 0.1736,941 ± 1,144
AdaIR28.828.81.211 ± 0.1996,828 ± 1,132
R2R29.629.60.118 ± 0.0671,056 ± 135
Ours (restoration)1691.82.00.863 ± 0.14511,861 ± 882
Ours (restoration + color mapping)1694.24.41.780 ± 0.41311,945 ± 565

Q3 Sensitivity to Clean/Uniformly Degraded Inputs & W1

In Eq. (3), the geometric inconsistency map U(q)U(q) is normalized via per-image min–max scaling onto [0,1][0,1]. Consequently, even if a negative has near-zero degradation—or uniform global degradation across the entire B channel—UU is forced to span the full [0,1][0,1] range, artificially amplifying residual updates in relatively clean regions. Furthermore, Eq. (1) and the architectural design strictly assume R and G are intact and only predict B^\hat{B}; in real archival films where G (magenta layer) or R (cyan layer) also suffer localized defects or scratches, the reference feature DRGD_{RG} will invert the guidance and corrupt the restoration.

Because NmmN_{\mathrm{mm}} in Eq.~(3) forces max⁡(U)=1\max(U)=1 and min⁡(U)=0\min(U)=0 for every image, how does the network behave when presented with an already-clean negative or a globally faded B-channel where relative spatial contrast in UU is pure noise?

Thank you for pointing out this limitation of our method. To evaluate its behavior on clean inputs, we directly fed the pseudo-ground-truth images, whose B channels are treated as undegraded, into our restoration network. For the globally faded case, we simulated global B-channel degradation by multiplying the B channel of each pseudo-ground-truth image by a constant attenuation factor α∈{0.9,0.5}\alpha \in \{0.9, 0.5\}, while keeping the R and G channels unchanged. We then evaluated these degraded images using our restoration network.

As shown in the following table, our method achieves reasonably good performance under both conditions, demonstrating robustness to clean inputs and uniform degradation, despite some performance degradation under stronger attenuation. We believe that incorporating more such cases into the training data could further improve the model’s performance.

MethodBlue Channel LPIPS↓Blue Channel PSNR↑Blue Channel SSIM↑Blue Channel FID↓Overall LPIPS↓Overall PSNR↑Overall SSIM↑Overall FID↓
Clean input0.019234.65000.989712.91000.010439.42000.99663.1000
Ours (α=0.9\alpha=0.9)0.022127.96060.987712.76710.019332.73180.99595.8907
Ours (α=0.5\alpha=0.5)0.032127.13550.980420.40730.029831.90670.993514.8178

 

Q4 Cross-Scene Reference Discrepancies

In Stage 2, when the reference print from the same film roll depicts a completely different scene with minimal semantic/geometric overlap (e.g., an indoor portrait vs. an outdoor landscape), does the softmax affinity matrix MM (Eq.5) with fixed temperature τ\tau genuinely become diffuse, or does it still peak on arbitrary false-positive patches? What value of τ\tau is used?

We use a fixed temperature of τ=0.04\tau=0.04 in the softmax affinity computation.

We compare 15 same-roll pairs with similar scenes and 15 same-roll pairs depicting substantially different scenes. We report the normalized entropy Hˉ\bar H, the average maximum affinity pˉmax⁡\bar p_{\max}, and the average top-3 reference mass Cˉ3\bar C_3. For query patch ii, AijA_{ij} denotes its affinity to reference patch jj, and Tk(i)T_k(i) denotes the indices of the kk largest affinities in row ii:

We summarize these quantities over all NtgtN_{\mathrm{tgt}} query patches as

\bar H &= \frac{1}{N_{\mathrm{tgt}}} \sum_{i=1}^{N_{\mathrm{tgt}}} \frac{H_i}{\log N_{\mathrm{ref}}}, \\ \bar C_3 &= \frac{1}{N_{\mathrm{tgt}}} \sum_{i=1}^{N_{\mathrm{tgt}}} C_i^{(3)}, \\ \bar p_{\max} &= \frac{1}{N_{\mathrm{tgt}}} \sum_{i=1}^{N_{\mathrm{tgt}}} p_i^{\max}. \end{aligned}$$ where $\varepsilon=10^{-8}$. The results are shown in below. | Pair type | $\bar H$ | $\bar p_{\max}$ | $\bar C_3$ | | --- | --- | --- | --- | | Same-roll, similar scenes | 0.56 | 0.15 | 0.26 | | Same-roll, different scenes | 0.73 | 0.06 | 0.12 | Compared with similar-scene pairs, different-scene pairs have higher entropy, lower largest affinity, and lower top-3 affinity mass. Thus, their affinities are relatively more dispersed across reference patches.

Follow-up questions on the proposed pipeline's viability for real archival film digitization:

The BlueNeg (Liu et al., 2025) benchmark provides 16-bit RAW film negative scans in Adobe DNG format, alongside 16-bit TIFF prints, and explicit film-stock metadata. Purpose of this was to avoid the severe dynamic-range clipping, quantization, and nonlinear gamma/inversion distortions introduced by 8-bit display-referred color space (e.g., sRGB or DCI-P3) conversions.

The proposed pipeline appears to discard the 16-bit RAW DNG negatives entirely and operate solely on BlueNeg’s downsampled 8-bit PNG negated previews.

  • Can the authors explicitly clarify whether all training and evaluation are conducted on the 8-bit negated PNG previews rather than the 16-bit RAW DNG negatives and 16-bit TIFF prints?

  • Since chemical dye fading and photochemical printing physically follow logarithmic optical density and linear transmittance rather than 8-bit display-referred color space (e.g., sRGB or DCI-P3), why do you apply synthetic B-channel fading augmentation (Appendix A.2) and learn Stage 2 print color mapping (Sec. 3.3) in 8-bit display space? Does reformulating them in 16-bit linear or log-density space improve real-fading generalization, highlight roll-off, and shadow chromaticity?

  • Can the pipeline process 16-bit RAW DNG negatives at the original resolution, or does relying on frozen DA3 features (pretrained on 8-bit display-referred color space (e.g., sRGB or DCI-P3) positive images) restrict your method to low-resolution, pre-inverted 8-bit previews? Providing an experiment/discussion on using the 16-bit RAW DNG data as input (e.g., operating in linear/log-density space while feeding tone-mapped views only to the frozen DA3 guidance branch) would clarify.

Reviewer_4rqgOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper focus on removing the channel-heterogeneous deterioration, especially for the blue (B) channel. This paper finds out that existing methods rarely deal with complementary cross-channel information. Therefore, the authors proposed a geometry-guided approach based on DA3. They further use the geometric features to guide reference-conditioned print color mapping by associating geometrically and semantically related regions across the restored negative and a reference printed photo.

Critical strengths
  1. I think the ablation design is careful. The guidance study steps from no geometric guidance to DA3 guidance (Table 1, bottom block) with nearly consistent performance gains. Also, the problem setting looks specific and useful to me. This paper proposes a method that removes channel-heterogeneous negative degradation with a geometry-guided approach. This task can have archival value, and the compared baselines were not built for it.
Critical weaknesses
  1. As far as I checked, every number in Tables 1 and 2 comes from a single training run with no variance estimate, while the margins that decide the rankings are only 0.31 to 0.70 dB. In my experience, gaps of this size could be simply reversed by a different random seed. Since both the superiority claims and the attribution argument are built on these margins, I think repeated runs (i.e., multi-seeds experiments, at least 3-seeds) with variance reported are needed to support them.
Questions for authors
  1. Could you describe more explicitly how DA3 and DINOv2 differ as used in the ablation? For example, is there a structural difference such that the three aligned channel views receive shared context (cross-view attention) in one case but are processed independently in the other?
  2. Could you report mean and variance over at least three training runs for the four rows that decide the rankings (i.e., Ours, R2R, DINO Guidance, and Uniform Affinity)?
  3. How was the BlueNeg pseudo ground truth constructed, and does its construction itself use cross-channel geometric structure?
2 comments

Thank you for the positive and valuable feedback. We address the concerns below and refer to the revised PDF, where edits are marked with blue text:

 

W1 & Q2 random seed experiment

As far as I checked, every number in Tables 1 and 2 comes from a single training run with no variance estimate, while the margins that decide the rankings are only 0.31 to 0.70 dB. In my experience, gaps of this size could be simply reversed by a different random seed. Since both the superiority claims and the attribution argument are built on these margins, I think repeated runs (i.e., multi-seeds experiments, at least 3-seeds) with variance reported are needed to support them.

Could you report mean and variance over at least three training runs for the four rows that decide the rankings (i.e., Ours, R2R, DINO Guidance, and Uniform Affinity)?

We repeated training for Ours, R2R, DINO Guidance, and Uniform Affinity using three independent random seeds per configuration and evaluated on the same test set. The results are reported in the following table, with each entry presented as mean / variance. Our method consistently outperforms the compared variants across the three runs, supporting the stability of the reported rankings. We will add the results in the Appendix D.2.

For restoration:

MethodBlue LPIPS↓Blue PSNR↑Blue SSIM↑Blue FID↓Overall LPIPS↓Overall PSNR↑Overall SSIM↑Overall FID↓
Ours0.0295 / 2.73×10⁻⁸29.4740 / 0.02060.9834 / 6.52×10⁻⁸20.7695 / 0.14230.0210 / 6.94×10⁻⁸34.2452 / 0.02060.9945 / 7.25×10⁻⁹8.6021 / 0.0114
DINO Guidance0.0309 / 5.74×10⁻⁹28.9255 / 0.00410.9827 / 8.18×10⁻⁹22.4124 / 0.12730.0224 / 2.33×10⁻⁷33.6967 / 0.00410.9942 / 9.09×10⁻¹⁰9.2585 / 0.1323
R2R0.0325 / 9.46×10⁻⁷28.7558 / 0.00790.9804 / 1.21×10⁻⁷21.5799 / 0.69810.0237 / 1.53×10⁻⁷33.5270 / 0.00790.9935 / 1.35×10⁻⁸10.4464 / 0.2399

For color mapping:

MethodSSIM↑PSNR↑LPIPS↓ΔE↓H-Corr↑H-Chi↓
Ours0.8894 / 2.29×10⁻⁷24.2945 / 2.85×10⁻³0.1968 / 7.31×10⁻⁷9.5761 / 0.04280.5876 / 3.86×10⁻⁴0.2585 / 9.54×10⁻⁵
Uniform Affinity0.8875 / 1.04×10⁻⁷23.9441 / 7.02×10⁻³0.2006 / 1.77×10⁻⁶9.9132 / 0.04680.5727 / 8.62×10⁻⁴0.2685 / 9.79×10⁻⁵

 

Q1 DINO and DA3 setting differences

Could you describe more explicitly how DA3 and DINOv2 differ as used in the ablation? For example, is there a structural difference such that the three aligned channel views receive shared context (cross-view attention) in one case but are processed independently in the other?

Yes. In our ablation, DA3 treats the three aligned channel views as three multi-view images, enabling cross-view attention to exchange information and distinguish differences in the underlying geometric structure across views. In contrast, DINOv2 itself does not support multi-view cross-attention; it processes each channel as an independent view, without shared context. This also explains our motivation for choosing DA3, as it provides geometry-aware priors and multi-view correspondences.

 

Q3 Pseudo GT implementation

How was the BlueNeg pseudo ground truth constructed, and does its construction itself use cross-channel geometric structure?

The BlueNeg pseudo ground truth is estimated via reverse printing. Specifically, a mapping is learned using blue-intact negatives and their corresponding prints to predict the negative's B channel:

n^b=f(prho,pgho,pbho,nr,ng)\hat{n}_b = f(p_r^{ho}, p_g^{ho}, p_b^{ho}, n_r, n_g)

where prhop_{r}^{\mathrm{ho}}, pghop_{g}^{\mathrm{ho}}, and pbhop_{b}^{\mathrm{ho}} are the RGB channels of the corresponding print, which may exhibit channel-homogeneous fading, and nrn_{r} and ngn_{g} are the negative's R and G channels. The fitted mapping is then applied to blue-corrupted samples to estimate their pseudo ground-truth B channels.

To our understanding, this mapping learns the photometric relationship between prints and negatives across image pairs, without explicitly modeling or comparing cross-channel geometric structure. In contrast, our method identifies deterioration-induced geometric discrepancies between color channels within each individual image. Thus, although both approaches use cross-channel information, they serve fundamentally different purposes.

Thank you for the response. After reading it, I have decided to maintain my score. However, I consider the inclusion of a non-anonymous GitHub link in the paper to be an ethical concern, as it may violate the double-blind review policy. I leave this matter to the judgment of the AC, SAC, and PC.

Reviewer_UJ3xOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

The paper addresses restoration of deteriorated color-film negatives, focusing on channel-heterogeneous degradation where the blue channel is often more corrupted than the red and green channels. The paper aims to exploit the fact that different color channels depict the same scene geometry: intermediate features from a frozen Depth Anything 3 (DA3) model are compared across channels to construct a spatial inconsistency map, which modulates restoration updates in a NAFNet-based network. The paper further uses DA3 features to establish correspondences between the restored negative and a reference printed photograph for spatially adaptive print-color reconstruction. Experiments on the BlueNeg dataset report improvements over general and all-in-one restoration methods, as well as existing reference-based color-mapping approaches.

Critical strengths
  • The paper identifies a meaningful structure specific to negative-film restoration: relatively preserved channels can provide complementary information for restoring a more severely corrupted channel. Using cross-channel feature inconsistency as spatial restoration guidance is intuitive and well-motivated.
  • The empirical results are promising. The proposed DA3-guided model improves over strong restoration baselines such as R2R, while the ablations comparing no guidance, pixel guidance, DINO guidance, and DA3 guidance provide useful evidence that the proposed guidance mechanism contributes to performance.
  • The use of a common pretrained representation for both restoration guidance and reference-conditioned color mapping provides a conceptually coherent two-stage framework. The color-mapping results also show improvements over several existing reference-based approaches.
Critical weaknesses
  • Limited evaluation and reliance on pseudo-ground truth
    • The restoration experiments are conducted entirely on BlueNeg, with training based on degraded-negative/pseudo-ground-truth pairs and additional synthetically generated B-channel fading. Importantly, the original BlueNeg paper notes that its pseudo ground truth is not true ground truth and may make the resulting evaluation unreliable, and therefore proposes an additional evaluation in the printed-photo domain; however, corresponding second restoration evaluation does not seem to be available in this paper.
  • The role of “geometry” is not sufficiently isolated
    • The paper attributes the improvement to geometry-aware information encoded by DA3, but the DA3-versus-DINO comparison does not fully isolate this factor: DA3 differs from DINO not only in geometric pretraining but also in how multiple channel views interact. The paper also notes that DA3 jointly processes the three channel views through cross-view attention. Although DINO guidance performs worse, this does not by itself demonstrate that geometry is the source of the improvement. Comparisons with another geometric foundation model or controlled DA3 variants would make this interpretation more convincing.
  • Strong assumption that degradation is concentrated in the B channel
    • The restoration network predicts only the B channel and directly preserves the degraded input's R/G channels. The reviewer understands that this is appropriate for the dominant degradation pattern considered in BlueNeg, but practical negatives may contain degradation in multiple channels or cases where R/G are not reliable structural references. The paper does not evaluate robustness when these assumptions are violated, which limits the demonstrated applicability of the proposed cross-channel strategy.
  • Incomplete comparison with recent image-restoration and color-grading methods
    • The paper includes several recent baselines, including RDBM/R2R and ColorFM/CanonCGT, but some additional recent relevant methods are not discussed and not compared against. In particular, UniRestorer [A], RestoreVAR [B], and FAPE-IR [C] are recent universal/all-in-one image-restoration approaches based respectively on adaptive multi-granularity expert selection, visual autoregressive generation, and frequency-aware MLLM-guided restoration. They are not specifically designed for negative-film degradation, so they are not necessarily direct apples-to-apples competitors; nevertheless, evaluating them on BlueNeg under the same retraining protocol, where feasible, would help establish whether the proposed negative-specific design provides benefits beyond recent general restoration frameworks. Furthermore, AceTone [D] is also relevant because it supports reference-image-conditioned color grading by generating 3D-LUT transformations, although it addresses color grading rather than negative restoration itself.

[A] Lin et al., UniRestorer: Universal Image Restoration via Adaptively Estimating Image Degradation at Proper Granularity. ICLR 2026.

[B] Rajagopalan et al., RestoreVAR: Visual Autoregressive Generation for All-in-One Image Restoration. ICLR 2026.

[C] Liu et al., FAPE-IR: Frequency-Aware Planning and Execution Framework for All-in-One Image Restoration. CVPR 2026.

[D] Ma et al., AceTone: Bridging Words and Colors for Conditional Image Grading. CVPR 2026.

Questions for authors
  • Could the authors report additional evaluation that reduces dependence on BlueNeg's pseudo-ground-truth restoration targets, for example through printed-photo-domain evaluation or testing on negatives from different acquisition/degradation conditions?
  • Can the authors provide stronger evidence that the advantage of DA3 arises specifically from geometric representations rather than differences in architecture, pretraining, or cross-view interaction?
  • How does the proposed method behave when the R or G channels are also significantly degraded? Can the cross-channel guidance mechanism identify which channels are reliable rather than assuming that only B requires restoration?
  • How sensitive is the color-mapping stage to reference selection? In particular, what happens when the reference comes from a different film roll or film stock, or when there is weak semantic/geometric correspondence between the reference and target?
  • Could the authors evaluate recent universal/all-in-one restoration approaches such as UniRestorer [A], RestoreVAR [B], or FAPE-IR [C] under the same BlueNeg training protocol, where applicable? For reference-conditioned color mapping, comparison with AceTone [D] would also help assess the proposed correspondence-based mapping against a recent reference-guided color-grading method.

For more information, please refer to the weaknesses section above.

6 comments

We thank for reviewer UJ3x's detailed responses. Below, we address the concerns (edits to the PDF are highlighted with blue text):

 

W1. Limited evaluation and reliance on pseudo-ground truth & Q1

The restoration experiments are conducted entirely on BlueNeg, with training based on degraded-negative/pseudo-ground-truth pairs and additional synthetically generated B-channel fading. Importantly, the original BlueNeg paper notes that its pseudo ground truth is not true ground truth and may make the resulting evaluation unreliable, and therefore proposes an additional evaluation in the printed-photo domain; however, corresponding second restoration evaluation does not seem to be available in this paper.

Could the authors report additional evaluation that reduces dependence on BlueNeg's pseudo-ground-truth restoration targets, for example through printed-photo-domain evaluation or testing on negatives from different acquisition/degradation conditions?

 

Dataset Availability and Evaluation Scope:

Our evaluation is limited to BlueNeg due to the lack of other suitable datasets for blue-channel restoration of deteriorated film negatives. UIEB [A] and Sea-Thru [B] contain digital photographs rather than film negatives and do not capture the same channel-dependent deterioration. The Prokudin-Gorskii Collection [C] lacks paired restoration targets for quantitative evaluation. In contrast, BlueNeg provides deteriorated negatives, corresponding prints, and derived pseudo-GT. To our knowledge, no other publicly available dataset offers comparable data and evaluation targets for this specific task.

Pseudo-GT reliability:

Since the original undamaged negatives are unavailable, exact ground truth cannot be obtained. Nevertheless, the pseudo-GT derived from the corresponding prints and R & G channel from negatives provides the closest available restoration targets for this setting. We therefore adopt them for quantitative evaluation, while acknowledging that they are approximations rather than exact ground truth.

Additional printed-domain analysis:

We additionally evaluate the results in the printed-photo domain. Following the histogram-matching protocol described in BlueNeg's printed-domain evaluation, we map the restored outputs into the color domain of the corresponding prints. As shown in following table, our method remains consistently superior to the competing methods across the reported metrics. This provides a complementary comparison in the printed color domain, although the reference still depends on pseudo-GT. We will add this additional experiment in Appendix D.3.

MethodBlue LPIPS↓Blue PSNR↑Blue SSIM↑Blue FID↓Overall LPIPS↓Overall PSNR↑Overall SSIM↑Overall FID↓
Ours0.055433.25350.973623.08460.021138.02470.99128.7337
R2R0.063632.06890.968925.35360.024636.84010.989611.1426
NAFNet0.065731.74320.967325.67590.026636.51440.98919.2774
AdaIR0.069631.70160.966026.79460.026636.47290.988712.9934
Restormer0.068931.65990.966328.46330.027436.43110.988811.7568
PromptIR0.118230.63060.923439.75650.036235.40190.974513.1064
AirNet0.093730.53850.952439.17340.037635.30970.984114.2246
MambaIRv20.086230.45040.958032.11650.032435.22170.986015.2718
UniRestore0.147927.30750.780647.99250.069232.07870.926920.9169

References

[A] C. Li, C. Guo, W. Ren, R. Cong, J. Hou, S. Kwong, and D. Tao, “An underwater image enhancement benchmark dataset and beyond,” IEEE Transactions on Image Processing, vol. 29, pp. 4376–4389, 2020.

[B] D. Akkaynak and T. Treibitz, “Sea-Thru: A method for removing water from underwater images,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 1682–1691, 2019.

[C] Library of Congress, “Prokudin-Gorskii Collection.” Available: https://www.loc.gov/collections/prokudin-gorskii/. Accessed: November 12, 2026.

W2 The role of “geometry” is not sufficiently isolated & Q2

The paper attributes the improvement to geometry-aware information encoded by DA3, but the DA3-versus-DINO comparison does not fully isolate this factor: DA3 differs from DINO not only in geometric pretraining but also in how multiple channel views interact. The paper also notes that DA3 jointly processes the three channel views through cross-view attention. Although DINO guidance performs worse, this does not by itself demonstrate that geometry is the source of the improvement. Comparisons with another geometric foundation model or controlled DA3 variants would make this interpretation more convincing.

Can the authors provide stronger evidence that the advantage of DA3 arises specifically from geometric representations rather than differences in architecture, pretraining, or cross-view interaction?

 

Our use of “geometry-aware” refers to exploiting structural relationships across the three channel views. Heterogeneous deterioration introduces discrepancies in the geometric structure inferred from individual channels, and these discrepancies are precisely what we seek to exploit. Cross-view attention provides a mechanism for identifying structural correspondences and disagreements across channels, allowing information from less deteriorated channels to guide restoration. We therefore regard this interaction as integral to the geometric guidance provided by DA3.

To disentangle the contributions of DA3 features and cross-channel interaction, we retrained our model with the RR, GG, and BB views processed independently through DA3, following the same setting as DINO, without cross-view attention between channels.

Additionally, we evaluated another geometry-pretrained model, VGGT [D], using features from its 9th layer (L9). The overall results are shown below. Both the separate-view DA3 variant and VGGT achieve higher PSNR than DINO on both the B channel and the overall RGB image, highlighting the importance of geometric priors in identifying cross-channel structural inconsistencies. Joint-view DA3 further improves performance, demonstrating the complementary benefits of cross-view attention for restoration.

MethodBlue Channel PSNR↑Overall PSNR↑
VGGT (L9)29.345834.1170
Ours (w.o. cross-attention)29.267034.0382
Ours29.560534.3317

Reference

[D] J. Wang, M. Chen, N. Karaev, A. Vedaldi, C. Rupprecht, and D. Novotny, “VGGT: Visual Geometry Grounded Transformer,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 5294–5306, 2025.

 

W3. Strong assumption that degradation is concentrated in the B channel & Q3

The restoration network predicts only the B channel and directly preserves the degraded input's R/G channels. The reviewer understands that this is appropriate for the dominant degradation pattern considered in BlueNeg, but practical negatives may contain degradation in multiple channels or cases where R/G are not reliable structural references. The paper does not evaluate robustness when these assumptions are violated, which limits the demonstrated applicability of the proposed cross-channel strategy. How does the proposed method behave when the R or G channels are also significantly degraded? Can the cross-channel guidance mechanism identify which channels are reliable rather than assuming that only B requires restoration?

 

BlueNeg contains images with damage confined to the red channel (e.g. ``19880215B-32-tama-safari-bogor.preview.png''), as well as those with blue-channel damage. Blue-channel deterioration predominates because the blue-sensitive layer is the outermost layer on the emulsion side, making it particularly vulnerable to environmental, chemical and physical damages. The dominance of blue-channel deterioration is a commonly observed phenomenon rather than an overly restrictive modelling assumption.

Our current implementation assumes that R/G channels provide reliable structural guidance for B-channel restoration, rather than explicitly estimating channel reliability. When R/G are also significantly degraded, this guidance may become unreliable. While the same cross-channel principle could be adapted to restore other individual channels through retraining, handling simultaneous multi-channel deterioration would require additional reliability estimation and suitable training data, which is beyond the scope of this work.

W4 Incomplete comparison with recent image-restoration and color-grading methods & Q5

The paper includes several recent baselines,..., but some additional recent relevant methods are not discussed and not compared against.

Could the authors evaluate recent universal/all-in-one restoration approaches such as UniRestorer [A], RestoreVAR [B], or FAPE-IR [C] under the same BlueNeg training protocol, where applicable? For reference-conditioned color mapping, comparison with AceTone [D] would also help assess the proposed correspondence-based mapping against a recent reference-guided color-grading method.

 

Additional restoration baselines.

We are retraining RestoreVAR on the BlueNeg training set. For FAPE-IR, training from scratch exceeds our available computational resources since the original training used 8 H200 GPUs. We therefore initialize from its released checkpoint and fine-tune it on BlueNeg. We also examined UniRestorer, but its public release does not provide the complete procedure needed to reproduce its multi-stage degradation clustering, expert training, and routing on a new dataset. We therefore cannot make a reliable BlueNeg retraining comparison with UniRestorer.

MethodBlue LPIPS↓Blue PSNR↑Blue SSIM↑Blue FID↓Overall LPIPS↓Overall PSNR↑Overall SSIM↑Overall FID↓
RestoreVAR0.474017.21310.5669263.53370.217821.98430.855661.8585
FAPE-IR0.226024.59330.836840.75210.226827.21360.846129.2084

Reference-conditioned color mapping.

For AceTone, since our BlueNeg dataset does not provide the extra LUT data and LUT-token targets required to retrain AceTone, we run its released checkpoint on the same color-mapping test set following the evaluation protocol for methods without compatible supervision in Sec 4.2 and Table 2. The results are:

MethodSSIM↑PSNR↑LPIPS↓ΔE\Delta E↓H-Corr↑H-Chi↓
AceTone0.727614.10900.382729.9800-0.01600.7322

AceTone performs below our correspondence-based color-mapping method on the reported metrics.

 

Q4 Reference-selection sensitivity

How sensitive is the color-mapping stage to reference selection? In particular, what happens when the reference comes from a different film roll or film stock, or when there is weak semantic/geometric correspondence between the reference and target?  

Our method does not necessarily require the reference and target to come from the same roll. When geometric correspondence between them is weak, the patch-wise affinity matrix in Eq. (5) tends to become diffuser, allowing the color mapping to rely more on global color information rather than localized correspondences. Here, we provide two settings to demonstrate this: different rolls with similar scene types, and the same roll but with completely different scenes (e.g., crowd and coastal scenery). We compare our method with baselines in paper for color mapping, with the results for each setting shown below.

As shown in the following table, our method achieves the best results on most of the metrics when using a reference from a different film roll with similar scene content. With a same-roll reference depicting a different scene, it still achieves the best PSNR, ΔE\Delta E, H-Corr, and H-Chi scores despite limited geometric correspondence. These results indicate that our method performs well under both reference settings, while references with similar scene content remain preferable for fully exploiting geometric correspondences.

Different roll, similar scene

MethodSSIM↑PSNR↑LPIPS↓ΔE↓H-Corr↑H-Chi↓
PhotoWCT20.748914.14860.433027.45890.01970.7904
CAP-VSTNet0.784819.31820.284316.82100.27250.4586
SA-LUT0.788619.75390.274014.58100.32200.3867
ColorFM0.832621.55340.277513.18400.22730.4870
Deep Preset0.888723.96190.202810.28550.53550.2472
CanonCGT0.832017.52750.310621.3780-0.03190.7431
Ours (DA3-guided)0.897624.70700.20009.59210.54360.2554

Same roll, different scene

MethodSSIM↑PSNR↑LPIPS↓ΔE↓H-Corr↑H-Chi↓
PhotoWCT20.768215.05100.344522.89840.20740.6594
CAP-VSTNet0.825721.45640.211212.37670.48360.2821
SA-LUT0.781019.32790.228715.23020.46120.2822
ColorFM0.830721.88690.214011.39200.52670.2600
Deep Preset0.874023.84270.205610.83170.58670.2661
CanonCGT0.810418.32140.286520.06470.07400.6585
Ours (DA3-guided)0.871324.21900.209510.19290.58850.2376

Comments on the paper

1 comment

We'd like to take this last chance to thank all reviewers again for thoughtful comments. We have made the following key updates (marked in blue in the rebuttal version) during the rebuttal period:

  • We added the NAFNet baseline under identical training settings to Table 1.

  • Explicitly stated that our main experiments use 8-bit previews due to the lack of full-resolution HDR pseudo-ground truth, and discussed extension to 16-bit RAW/DNG inputs as future work.

  • We added following supplementary experiments in the appendix: computational complexity comparisons (Appendix D.1), three-run stability analysis (Appendix D.2), printed-domain evaluation (Appendix D.3), and preliminary experiments with DNG-derived inputs, including qualitative results (Appendix D.4).

  • We expanded the appendix to clarify the roles of synthetic B-channel fading as spatial degradation augmentation (Appendix A.2).

  • I sincerely apologize for inadvertently including a non-anonymous GitHub link to share our codes in the original submission. As soon as we became aware of the issue, I made the repository private. Now the link from the revised manuscript has been replaced with an anonymized code link.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.