{"id":"6d5f07e7-0d91-4e4c-974e-733cc7c1ce65","arxiv_id":"2411.13150","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"RAW-Diffusion generates high-fidelity RAW images from RGB inputs with state-of-the-art PSNR/SSIM on four DSLR datasets, and needs as few as 25 training images.","lead":"This paper presents a diffusion model that reconstructs camera RAW images from ordinary RGB photos, guided by features extracted from the RGB input. If it holds up, it could let researchers create sensor-specific RAW training data from existing RGB datasets, reducing the cost of data collection for low-light computer vision.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Training-protocol confound in Section 4.4 is acknowledged by the authors and load-bearing for the data-efficiency claim; the RGB2RAW SOTA claim is not threatened. A direct retraining check would settle it.","rationale":"The reader's weakest_assumption identifies the same confound I consider load-bearing: Section 4.4 mixes epoch-based full training with iteration-based few-shot training, and the paper itself states the iteration-based schedule improves performance. My independent reading of the paper confirms the numerical pattern (full-training PSNR is sometimes below several few-shot runs on the same dataset, e.g., FiveK Nikon 25 images 31.29 vs full 30.05; FiveK Canon 250 images 35.93 vs full 34.01) and that no control experiment separates the protocol variable from the sample-count variable. That is the most defensible route to a load-bearing concern. I also weigh the positive evidence: the RGB2RAW SOTA comparison in Table 1 is a fair, same-protocol comparison and the gains over InvISP+/ISPLess+ are large; the conditioning ablations (Tables 6-10) support the architectural choices; DDIM sampling at 6 steps is a useful independent robustness result; and the downstream detection gains (Tables 3-5) hold under a fixed training protocol (100 NOD images plus generated data at pgen=0.95), so they are not contaminated by the Section 4.4 schedule issue. The promised code at the GitHub URL is a checkable artifact, but I did not verify it; that is a reproducibility risk rather than an argument-level flaw. My verdict remains CONDITIONAL with the condition being the protocol-controlled retraining experiment described in concrete_test. If that experiment shows the parity holds, the paper could be accepted; if not, the headline data-efficiency claim would need substantial softening. I do not see grounds for REJECT because the central reconstruction method is well supported and the confound is explicitly disclosed rather than hidden, but the disclosed confound is severe enough that unconditional acceptance is premature. I agree with the reader's assessment rather than partially disagreeing: the reader pinpointed the same confound, and I have not identified any additional load-bearing concern of comparable weight, such as an internal inconsistency in the method or an unsupported claim about the detector experiments.","tokens_in":22411,"tokens_out":2065,"duration_ms":18328,"concrete_test":"Retrain the full-data model with the iteration-based protocol used in Section 4.4 (same 70k total steps, same batch size, same optimizer, data repeated to fill epochs) and compare to the epoch-based full run in Table 2. Also train a 25-image model with the epoch-based protocol. If iteration-based full training matches or exceeds the epoch-based full run and the 25-image parity persists, the data-efficiency claim is supported. If full-data performance under the iteration-based protocol rises to the 25-image level (e.g., FiveK Nikon toward 31.29), the claimed parity disappears and Table 2 should be reinterpreted as a protocol effect. Report which specific 25/50/... NOD images are used so the subset is reproducible.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim most at risk is the headline data-efficiency result: 'achieving remarkable performance with as few as 25 training samples or even fewer.' Section 4.4 trains the full model epoch-based for 70k steps, while all subsampled models use an iteration-based schedule with the small sets 'repeated accordingly.' The authors state this iteration-based schedule 'improves performance,' so the small-data runs may be better because of the training protocol, not because the model is intrinsically sample-efficient. The effect is visible numerically: on FiveK Nikon, 25-image training yields PSNR 31.29 versus 30.05 for full training, and on FiveK Canon 250-image training yields 35.93 versus 34.01, so the protocol advantage appears large enough to mask or invert the true sample-size effect. The paper acknowledges this confound explicitly, so it is internally consistent but unexplained: no ablation varies the protocol while holding dataset size fixed. This concern is load-bearing for the data-efficiency headline because the main quantitative support (Table 2 parity at 25 images) is exactly the comparison between the epoch-based full run and the iteration-based few-shot runs. The RGB2RAW SOTA claim (Section 4.3, Table 1) and the detector-gain claim (Tables 3-5) are not affected by this confound. A second, weaker concern is that the NOD few-shot training subsets are not specified (which 25 of 3.2k/4.0k images), so subset selection is uncontrolled; this is secondary because Table 2 shows robustness across multiple sample sizes and datasets.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes RAW-Diffusion, a denoising-diffusion model that reconstructs RAW sensor images from RGB images. The method uses an EDSR-based RGB-guidance module and RGB-guided residual blocks (SPADE-style modulation) injected into a U-Net at multiple resolutions, trains with a sum of MSE, L1, and log-L1 losses to directly predict the denoised RAW image, and uses DDIM for fast sampling. Experiments on Four DSLR datasets (FiveK Nikon/Canon, NOD Nikon/Sony) report PSNR/SSIM gains over prior RGB2RAW methods, e.g., 30.05 dB on FiveK Nikon and 44.93 dB on NOD Nikon in Table 1. The paper further claims that training on as few as 25 images matches or exceeds full-data performance (Table 2), and shows that detectors trained with BDD100K-RAW generated by RAW-Diffusion outperform those trained with SRISP-generated RAW (Tables 3-5).","tokens_in":22763,"tokens_out":4282,"duration_ms":40335,"significance":"If the results hold, the paper makes a useful methodological contribution: it is, to my knowledge, the first diffusion-based RGB2RAW method to outperform inversion-based and ISP-simulation baselines on these benchmarks, and the downstream detection experiment gives a practical demonstration of generating sensor-specific RAW datasets from existing RGB corpora. The contribution is supported by careful ablations of prediction parameterization, conditioning architecture, losses, backbones, and sampling steps, and the code is released. The main caveat concerns the data-efficiency claim, which is not yet supported by a controlled comparison; with that resolved, the paper would be a solid addition.","major_comments":[{"comment":"The data-efficiency claim is confounded by the training protocol. The full model is trained with an epoch-based schedule for 70k steps, while all subsampled models are trained with an iteration-based schedule in which the small datasets are repeated; the authors explicitly state that the iteration-based schedule 'improves performance.' Table 2 shows that 25-image training on FiveK Nikon achieves 31.29 dB PSNR versus 30.05 dB for full training, and 250-image training on FiveK Canon achieves 35.93 versus 34.01, so the protocol advantage is large enough to mask the effect of sample size. The claim that 'RAW-Diffusion achieves the same performance with 25 or fewer training samples as when training on the full training set' requires a comparison in which the full model is retrained with the same iteration-based schedule (or the few-shot models are trained epoch-based). Without that control, the parity is evidence about the training schedule, not about sample efficiency. Because this is a headline contribution, I consider this load-bearing.","section":"§4.4, Table 2"},{"comment":"The NOD few-shot experiments do not specify how the subsampled training sets of 500, 250, 100, 50, 25, and 10 images are drawn from the 3.2k/4.0k NOD training images. For FiveK the split is fixed and follows [89], but for NOD no random-seed or selection procedure is reported, so subset composition is uncontrolled; a favorable subset could partly explain the flat or improving PSNR as training set size decreases. Please specify the selection protocol and, ideally, report results over multiple subset draws.","section":"§4.4"},{"comment":"Several baselines in Table 1 are evaluated under modified settings that are not applied uniformly. InvISP and ISPLess are retrained as 'InvISP+' and 'ISPLess+' with JPEG simulation removed and ground-truth RGB input; RISPNet is downscaled and run without ensembling; the diffusion baseline is trained on 64x64 patches; and SRISP uses the mean global feature as test-time reference. Some modifications are reasonable (e.g., avoiding JPEG artifacts in evaluation), and the reported margins are large, but the headline 'state-of-the-art' comparison is not fully uniform. I ask the authors to state clearly, for each baseline, which configuration is used and whether any modifications were needed to adapt it to the evaluation protocol, and to consider reporting results for the original configurations as well.","section":"§4.2–§4.3, Table 1"}],"minor_comments":[{"comment":"The NOD Sony, AP_S column reports '.8±0.3'; the leading zero appears to be missing and should read '0.8±0.3'.","section":"Table 14"},{"comment":"In the third paragraph, 'RSISP [63]' should be 'SRISP [63]'.","section":"§4.5"},{"comment":"The sentence 'When increasing the number of training runs, the results better align with the expectations' is unclear; please rephrase to say that results become more consistent as the number of random seeds increases.","section":"§4.4"},{"comment":"Table 2 reports no standard deviations, while Table 1 does; adding standard deviations or confidence intervals would help readers assess whether the small-data differences are significant.","section":"Table 2"},{"comment":"The text says 'To maintain a consistent number of training steps, the small training sets are repeated accordingly.' Please clarify whether the total number of optimizer steps is exactly 70k for all subsampled runs, and whether the learning-rate schedule is reset per run or aligned across runs.","section":"§4.4"},{"comment":"For the diffusion baseline, the statement 'Diffusion is trained on 64×64 patches as described by the authors. The performance with larger patch sizes collapses' would benefit from a brief explanation, since the patch size may disadvantage the baseline relative to the 256×256 patches used for the proposed method.","section":"§4.2"}],"recommendation":"major_revision","confidential_remarks":"The SRISP baseline is from the same research group (one Sony AI co-author), and it is used as a comparison baseline rather than as evidence. This is not a problem per se, but it would be prudent to disclose the relationship in the manuscript. The main evaluation concern is the epoch/iteration confound in Section 4.4; if the authors can supply a controlled comparison, the paper would be suitable. The fit to the journal's scope is good."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The main thing to know: this is a credible and useful RGB-to-RAW diffusion paper. The headline RGB2RAW SOTA claim holds up — Table 1 shows clear gains over InvISP+, SRISP, and the prior diffusion baseline on four DSLR datasets, and the ablations (direct RAW prediction vs. noise prediction, RGB-guided blocks vs. concatenation, loss composition) are clean and support the architecture choices. The downstream detection experiment is also well executed: generated BDD100K-RAW and Cityscapes-RAW improve AP over SRISP-generated data, and the zero-shot results are informative.\n\nWhat is actually new is the specific combination: RGB-guidance features injected via SPADE-style residual blocks at multiple scales, direct RAW prediction with a logL1 objective, and the demonstration that this works with very few training images. The 25-sample result is the headline, and here is where I have a real concern. The full model is trained epoch-based for 70k steps, while the subsampled models use an iteration-based schedule. The authors state in Section 4.4 that the iteration-based schedule improves performance. That is an explicit admission that the few-shot runs benefit from a better training protocol, not necessarily from sample efficiency. Numerically it matters: 25-image training beats full training by 1.24 dB on FiveK Nikon. The stress-test note is right that the data-efficiency claim is load-bearing on this confound. It does not threaten the SOTA claim or the detector-gain claim, but it does mean the \"as few as 25 samples\" headline is not yet established. The fix is straightforward: retrain the full model under the iteration-based schedule and see if parity holds.\n\nA secondary, minor issue: the NOD few-shot subsets are not specified — which 25 of 3.2k images? — so subset selection is uncontrolled. Table 2's robustness across sample sizes and datasets softens this, but it is still worth asking for.\n\nThe baseline modifications (InvISP+, ISPLess+) are disclosed and reasonable, and the comparison to the prior diffusion work [52] uses the authors' own patch size. The citation pattern is fine; SRISP is from the same group but used as a baseline, not as support. The paper ships code (with a caveat that it is a \"published code\" link, but the promise is there) and reports three-seed means.\n\nWho this is for: anyone working on RAW image synthesis, camera-specific dataset generation, or low-light perception. It deserves a serious referee. My recommendation: send to peer review, flag the protocol confound, and ask for the retrain check before the data-efficiency claim can be taken at face value.","headline":"Solid RGB2RAW engineering with a real data-efficiency claim that is currently undercut by a training-protocol confound, which the authors themselves acknowledge.","tokens_in":23300,"tokens_out":952,"would_cite":true,"duration_ms":11840,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"RAW-Diffusion is a diffusion model that reconstructs high-fidelity RAW images from RGB inputs, reports state-of-the-art performance on all four tested DSLR datasets, and matches full-data performance with only 25 training images.","keywords":["RAW image generation","RGB-to-RAW reconstruction","diffusion models","conditional image generation","data efficiency","object detection","reverse ISP","sensor-specific datasets"],"falsifier":"Retrain the full-data model with the same iteration-based schedule used for the 25-image runs and compare PSNR on the four test sets; if the full model clearly surpasses the 25-image model, the sample-efficiency parity claim fails. Alternatively, train a detector exclusively on RAW-Diffusion-generated RAW for a camera whose real RAW pairs were held out and check whether its AP stays within a few points of a detector trained on real RAW from that camera.","tokens_in":22224,"feed_emoji":"📷","tokens_out":9320,"duration_ms":81647,"temperature":0.7,"pith_summary":"RAW-Diffusion sets out to solve a practical bottleneck: RAW sensor images carry more information than RGB, but collecting and storing RAW datasets for each camera is expensive. The paper argues that a denoising diffusion model conditioned on an RGB image can reconstruct the corresponding RAW sensor values with higher fidelity than existing inverse-ISP networks, and that this conditioning makes the model extremely data-efficient, with 25 training images matching the full training set. If true, large existing RGB datasets could be converted into camera-specific RAW datasets, reducing the need for costly sensor-specific data collection. The paper supports this with state-of-the-art PSNR and SSIM on four DSLR datasets and with object-detection experiments using generated BDD100K-RAW and Cityscapes-RAW.","feed_headline":"RGB-to-RAW diffusion tops prior methods on four DSLRs","feed_subtitle":"Uses RGB-guided denoising to reconstruct sensor data, matching full-dataset quality with 25 training images","key_machinery":"The load-bearing mechanism is a conditional denoising U-Net with two novel insertion points. An RGB-guidance module, an EDSR-style stack of residual blocks without an upsampling head, extracts a guidance feature map $F_{RGB}$ from the input image. These features enter the reverse diffusion process through RGB-guided residual blocks that apply a spatially adaptive modulation to the normalized diffusion features: $F^{i+1} = \\mathrm{Norm}(F^i)\\cdot(1+\\gamma_i(\\mathrm{Down}(F_{RGB})))+\\beta_i(\\mathrm{Down}(F_{RGB}))$, where $\\gamma_i$ and $\\beta_i$ are predicted per-channel, per-pixel modulation parameters. The block is used at the bottleneck and every decoder level, so the RGB context shapes all resolutions of the reconstruction. The model directly predicts the RAW image rather than the noise, uses a hyperbolic tangent output, and is trained with a sum of MSE, L1, and logarithmic L1 losses to handle the skewed distribution of RAW values.","core_discovery":"The central claim is that a diffusion model can invert the image signal processing pipeline: given an RGB image, RAW-Diffusion iteratively denoises random noise into a RAW image whose pixel values match the original sensor readings, and it does this better than prior learned and calibration-based methods. On the test sets, the method reports PSNR of 30.05 (FiveK Nikon), 34.01 (FiveK Canon), 44.93 (NOD Nikon), and 39.17 (NOD Sony), the best among the eleven compared methods. The paper also claims that the same model trained on only 25 images performs on par with full-data training, and that RAW images generated from BDD100K and Cityscapes improve Faster R-CNN and YOLOv8 detectors more than the same datasets converted by SRISP, including zero-shot settings where the detector never sees real NOD images.","pith_inferences":["If the data-efficiency result extends to other sensors, the method could turn any large annotated RGB dataset into a RAW training set for low-light or edge vision, bypassing the per-sensor collection bottleneck that limits RAW-based computer vision.","A multi-sensor variant that conditions on camera identity is a natural next step; the paper lists sensor generalization as future work.","A strict comparison of the few-shot and full-data training protocols is needed before '25 images match the full dataset' becomes a design rule, since the paper notes that the iteration-based few-shot schedule improves performance.","The downstream gains from generated RAW may partly reflect distributional realism rather than pixel accuracy; measuring detector performance against PSNR at fixed sample budgets would separate these factors."],"forward_implications":["On all four DSLR test sets, reconstructed RAW images are closer to the ground-truth sensor output than those of all eleven compared methods, with the largest margin on the low-light NOD Nikon set.","Training with 25 images, and in some cases 10, yields PSNR within the full-data range, so a usable sensor-specific RAW generator could in principle be built from a very small paired capture session.","Converting large RGB datasets such as BDD100K and Cityscapes into RAW with RAW-Diffusion improves object detection on NOD more than the same conversion with SRISP, both when mixed with real NOD images and in zero-shot training.","DDIM sampling with as few as 6 steps matches 1000-step DDPM reconstructions, which makes large-scale dataset generation computationally practical.","Ablation results show direct RAW prediction, the RGB-guided residual blocks, and the logarithmic L1 loss each contribute; predicting noise instead collapses PSNR to 5.55 on NOD Nikon."],"supporting_citations":[{"why":"supplies the DDPM forward and reverse diffusion processes that the method is built on.","marker":"[32]"},{"why":"provides the EDSR residual-block backbone used in the RGB-guidance module.","marker":"[44]"},{"why":"introduces spatially-adaptive normalization, which the RGB-guided residual blocks adapt.","marker":"[64]"},{"why":"demonstrates SPADE-conditioned diffusion for semantic maps and inspires the guidance injection.","marker":"[83]"},{"why":"is the main inverse-ISP baseline and supplies the reconstruction evaluation protocol and test split.","marker":"[89]"},{"why":"is the recent RGB2RAW baseline used both for reconstruction comparison and for generating comparison RAW datasets.","marker":"[63]"},{"why":"provides DDIM sampling, used to reduce the 1000-step process to as few as 6 steps.","marker":"[78]"},{"why":"supplies the FiveK paired RGB-RAW images used for reconstruction experiments.","marker":"[6]"},{"why":"supplies the NOD low-light RAW dataset with object annotations used in downstream detection.","marker":"[58]"},{"why":"supplies BDD100K, the large-scale RGB dataset converted into RAW for detection training.","marker":"[93]"}],"fun_headline_variants":["Diffusion flips RGB to RAW, tops on four DSLR sets","RAW-Diffusion: 25 RGB images generate state-of-the-art RAW","RGB-to-RAW diffusion beats 11 methods on DSLR datasets","Diffusion model turns RGB into RAW with only 25 training shots","State-of-the-art RAW generation from RGB via diffusion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The 25-image sample-efficiency claim assumes that the iteration-based training schedule used for the small-data runs is comparable to the epoch-based schedule used for the full-data run, since the paper notes that the iteration-based schedule improves performance.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion flips RGB to RAW, tops on four DSLR sets","RAW-Diffusion: 25 RGB images generate state-of-the-art RAW","RGB-to-RAW diffusion beats 11 methods on DSLR datasets","Diffusion model turns RGB into RAW with only 25 training shots","State-of-the-art RAW generation from RGB via diffusion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000409,"raw_usage":{"total_tokens":2105,"prompt_tokens":915,"completion_tokens":1190,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":531,"completion_tokens_details":{"reasoning_tokens":1099}},"tokens_in":531,"tokens_out":1190,"duration_ms":9172,"temperature":1.0,"reasoning_tokens":1099,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T16:46:40.099316+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the full-data model with the same iteration-based schedule used for the 25-image runs and compare PSNR on the four test sets; if the full model clearly surpasses the 25-image model, the sample-efficiency parity claim fails. Alternatively, train a detector exclusively on RAW-Diffusion-generated RAW for a camera whose real RAW pairs were held out and check whether its AP stays within a few points of a detector trained on real RAW from that camera.","supporting_citations":[{"cited_title":"Denois- ing diffusion probabilistic models","cited_arxiv_id":null,"evidence_quote":"supplies the DDPM forward and reverse diffusion processes that the method is built on."},{"cited_title":"Enhanced deep resid- ual networks for single image super-resolution","cited_arxiv_id":null,"evidence_quote":"provides the EDSR residual-block backbone used in the RGB-guidance module."},{"cited_title":"Self-Supervised Reversed Image Signal Processing via Reference-Guided Dynamic Parameter Selection","cited_arxiv_id":"2303.13916","evidence_quote":"introduces spatially-adaptive normalization, which the RGB-guided residual blocks adapt."},{"cited_title":"Exploiting diffusion prior for real-world image super-resolution","cited_arxiv_id":null,"evidence_quote":"demonstrates SPADE-conditioned diffusion for semantic maps and inspires the guidance injection."},{"cited_title":"Smartbrush: Text and shape guided ob- ject inpainting with diffusion model","cited_arxiv_id":null,"evidence_quote":"is the main inverse-ISP baseline and supplies the reconstruction evaluation protocol and test split."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"is the recent RGB2RAW baseline used both for reconstruction comparison and for generating comparison RAW datasets."},{"cited_title":"Deep unsuper- vised learning using nonequilibrium thermodynam- ics","cited_arxiv_id":null,"evidence_quote":"provides DDIM sampling, used to reduce the 1000-step process to as few as 6 steps."},{"cited_title":"Radiomet- ric self calibration","cited_arxiv_id":null,"evidence_quote":"supplies the NOD low-light RAW dataset with object annotations used in downstream detection."},{"cited_title":"Rawgment: noise-accounted raw augmentation enables recognition in a wide variety of environments","cited_arxiv_id":null,"evidence_quote":"supplies BDD100K, the large-scale RGB dataset converted into RAW for detection training."}],"review_version":1}