- 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 , 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
- 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).
- 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
- Unaddressed Failure Mode in Per-Image Min–Max Normalization (): In Eq. (3), the geometric inconsistency map is normalized via per-image min–max scaling onto . Consequently, even if a negative has near-zero degradation—or uniform global degradation across the entire B channel— is forced to span the full 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 ; in real archival films where G (magenta layer) or R (cyan layer) also suffer localized defects or scratches, the reference feature will invert the guidance and corrupt the restoration.
- 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
- 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?
- 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?
- Sensitivity to Clean/Uniformly Degraded Inputs: Because in Eq. (3) forces and 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 is pure noise?
- 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 (Eq. 5) with fixed temperature genuinely become diffuse, or does it still peak on arbitrary false-positive patches? What value of is used?
11 comments
Q3 Sensitivity to Clean/Uniformly Degraded Inputs & W1
In Eq. (3), the geometric inconsistency map is normalized via per-image min–max scaling onto . Consequently, even if a negative has near-zero degradation—or uniform global degradation across the entire B channel— is forced to span the full 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 ; in real archival films where G (magenta layer) or R (cyan layer) also suffer localized defects or scratches, the reference feature will invert the guidance and corrupt the restoration.
Because in Eq.~(3) forces and 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 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 , 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.
| Method | Blue Channel LPIPS↓ | Blue Channel PSNR↑ | Blue Channel SSIM↑ | Blue Channel FID↓ | Overall LPIPS↓ | Overall PSNR↑ | Overall SSIM↑ | Overall FID↓ |
|---|---|---|---|---|---|---|---|---|
| Clean input | 0.0192 | 34.6500 | 0.9897 | 12.9100 | 0.0104 | 39.4200 | 0.9966 | 3.1000 |
| Ours () | 0.0221 | 27.9606 | 0.9877 | 12.7671 | 0.0193 | 32.7318 | 0.9959 | 5.8907 |
| Ours () | 0.0321 | 27.1355 | 0.9804 | 20.4073 | 0.0298 | 31.9067 | 0.9935 | 14.8178 |
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 (Eq.5) with fixed temperature genuinely become diffuse, or does it still peak on arbitrary false-positive patches? What value of is used?
We use a fixed temperature of 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 , the average maximum affinity , and the average top-3 reference mass . For query patch , denotes its affinity to reference patch , and denotes the indices of the largest affinities in row :
We summarize these quantities over all 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.
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.
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.