{"id":"df7ec6c8-2c1f-48a8-ace8-d9077fef9b12","arxiv_id":"2509.01837","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"PractiLight is a data-efficient relighting method that uses a lightweight LoRA regressor on stable diffusion self-attention layers to guide generation toward a target irradiance map.","lead":"A new relighting method trains a tiny LoRA adapter on a diffusion model's self-attention layers to predict lighting maps, then uses it to guide image generation toward a desired light condition. It aims to control light across many image styles using only a few thousand synthetic training examples.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Control-adherence metrics compare relit images to the direct-irradiance control map itself; a naive source/control blend beats PractiLight on these metrics, so the 'state-of-the-art control' claim is not numerically supported.","rationale":"I read the paper as claiming that a 790K-parameter LoRA regressor trained on 4K synthetic primitive scenes, used as classifier guidance, gives state-of-the-art relighting quality and control with large parameter/data/compute savings. For that claim to hold, the experiments must demonstrate both that the desired lighting actually appears in the output and that identity/style is preserved. Identity and quality are supported by a user study, HPS scores, and qualitative comparisons. Control, however, is not convincingly validated. The control metrics in Table 1 are computed directly between the output image and the control irradiance map I_c. That map is a rendered shading field, not a target photograph, so perceptual similarity to it is not synonymous with correct light transport. The authors' own Table 2 exposes this: a naive 50/50 blend of I_s and I_c scores better than the full method on every control metric, which means those metrics reward appearance similarity to the control map regardless of whether the result is a plausible relighting. The user study also does not ask the key control question—whether the lighting direction, shadow geometry, or highlight placement matches the condition. Therefore the 'state-of-the-art control' prong of the central claim is unsubstantiated by the published evidence. This reinforces the reader's CONDITIONAL verdict rather than overturning it: the method may still be effective, but the control claim needs a physically meaningful evaluation before it can be accepted. My proposed test—measuring estimated light-direction error against the known rendering condition—directly targets this gap and would settle whether the concern lands.","tokens_in":17152,"tokens_out":6586,"duration_ms":74480,"concrete_test":"Take a subset (e.g., 50 images) of the evaluation set where I_c was rendered from a known point-light position. Have independent annotators (or an automatic light-direction estimator) mark the dominant light direction and cast-shadow direction in each output of PractiLight, IC-Light, RGB↔x, and Blending. Compare angular error against the known direction used to render I_c. If PractiLight does not have significantly lower angular error than IC-Light, or if Blending does not fail clearly, the control-adherence claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Sec. 5.4's 'Control' metrics (L2/LPIPS/CLIP/DINO) are computed between the relit image and I_c, a rendered direct-irradiance map. I_c is not a picture of the relit scene; it is a shading field on a depth-proxy. A correct relit image should not be close to I_c in pixel or deep-feature space. This is not a minor point: in Table 2 the 'Blending' baseline (50/50 blend of I_s and I_c) beats the full method on all four control metrics (L2 0.057 vs 0.130; LPIPS 0.61 vs 0.74; CLIP 0.47 vs 0.56; DINO 0.59 vs 0.71) while being visually unacceptable. Thus the Table 1 claim that PractiLight is 'on par' on control adherence is not established by these numbers; on the published metrics it is actually worse than IC-Light on every control column (e.g., L2 0.13 vs 0.10, LPIPS 0.74 vs 0.63). The user study (Sec. 5.5, Table 3) only asks about identity/style and overall relighting preference; it never asks whether the target lighting direction, shadow placement, or highlight shape was achieved. Since 'quality and control' is the headline claim, the control half currently rests on a metric that a trivial baseline passes better than the method.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PractiLight, a relighting method that trains a lightweight LoRA regressor on the self-attention layers of Stable Diffusion 1.5 to predict direct-irradiance maps from images. The regressor is trained on only 4,000 synthetic images of simple geometric primitives, and is then used as a classifier-guidance energy term during sampling, together with edge-ControlNet and self-attention query injection, to relight generated images according to a user-provided lighting condition. The authors claim state-of-the-art quality and control with roughly 100x fewer parameters and 1000x less data than leading methods such as IC-Light and RGB↔x. The evidence includes quantitative metrics, user studies, ablations, and qualitative results across portraits, anime, sketches, paintings, and other domains.","tokens_in":17571,"tokens_out":4504,"duration_ms":49258,"significance":"If the central claims hold, PractiLight would be a practically valuable contribution: it is parameter- and data-efficient, taps into a frozen foundation model rather than fine-tuning it, and the analysis of self-attention layers as carriers of light-transport information is a useful insight for the community. The paper also ships a relatively thorough ablation study, including dataset-size effects and design choices. However, the current evaluation does not substantiate the headline claim of state-of-the-art control: the control metrics appear to measure similarity to a shading field rather than to a relit image, and a trivial blending baseline outperforms the full method on those metrics. The user study does not test whether the target lighting direction or shadow placement was achieved. These issues are load-bearing and require substantive revision.","major_comments":[{"comment":"The 'Control' metrics compare the relit image to I_c, the direct-irradiance map rendered on a depth-proxy geometry. I_c is not a picture of a relit scene; it is a shading field without albedo, texture, or full light transport. A correct relit image should not be close to I_c in pixel or deep-feature space. This is not a minor issue: Table 2 shows that the 'Blending' baseline (50/50 blend of I_s and I_c) beats the full method on all four control metrics (L2 0.057 vs 0.130; LPIPS 0.61 vs 0.74; CLIP 0.47 vs 0.56; DINO 0.59 vs 0.71) while being visually unacceptable. Moreover, in Table 1 the full method is worse than IC-Light on every control column (e.g., L2 0.13 vs 0.10; LPIPS 0.74 vs 0.63). The claim that PractiLight is 'on par' in control adherence is therefore not supported by the published numbers. Please redesign the control evaluation, for example using rendered ground-truth relit sc","section":"§5.4, Tables 1 and 2"},{"comment":"The subjective evaluation asks only two questions: which method better preserves identity/style, and which produces an overall better relighting result. It never asks whether the target lighting direction, shadow placement, highlight shape, or light intensity was achieved. Since the paper's headline claim is 'quality and control,' the control half of that claim is not validated by the user study. Please add direct control questions (e.g., 'which result better matches the lighting direction shown in the condition?') or a separate perceptual control study.","section":"§5.5, Table 3"},{"comment":"No error bars, confidence intervals, or statistical significance tests are reported. The HPSv2 differences between Ours (0.2493), IC-Light (0.2462), and RGB↔x (0.2334) may be within noise, especially with a self-constructed dataset of only 180 images. The user-study percentages in Table 3 also lack confidence intervals. Please report per-prompt variances, repeated-seed runs, or bootstrap intervals, and state whether the observed differences are significant.","section":"§5.4 and Table 1"},{"comment":"The evaluation dataset is self-constructed, and the control signals are generated using the paper's own pipeline (§4.2). Hyperparameters for all methods, including PractiLight, appear to be tuned on the same evaluation dataset: the text says CFG values and diffusion timesteps were chosen 'between dataset-provided values and manual tuning (the best of the two options)' and that hyperparameters were 'tuned and fixed ... to the best of our abilities.' This creates a risk of overfitting to the test set. Please use a separate validation split for hyperparameter selection and report results on a held-out test set, or at least report sensitivity to the chosen hyperparameters.","section":"§5.2 and §5.4"},{"comment":"The regressor is trained only on 4,000 synthetic images of spheres, cubes, and cuboids with a single point light source, yet it is used to guide relighting of portraits, anime, sketches, paintings, and complex scenes. The generalization of the regressor across this domain gap is a load-bearing assumption, but it is validated only qualitatively. Please provide quantitative evidence, such as regressor prediction error on held-out synthetic domains, or a correlation between regressor accuracy and relighting success on the evaluation dataset. Without this, the method's broad-domain applicability remains plausible but not established.","section":"§4.1 and §5"}],"minor_comments":[{"comment":"The figure uses timesteps t=900 and t=100, while later text defines t as a normalized timestep with t=0 as pure noise and t=1 as the final image. Please make the notation consistent or clarify the mapping.","section":"§3.1, Fig. 3"},{"comment":"The text says PractiLight is 'on par in terms of control adherence,' but Table 1 numerically shows it is worse than IC-Light on all four control metrics. Please rephrase to match the reported numbers.","section":"§5.4"},{"comment":"For DiLightNet comparisons, the supplement states that 10 different random light sources were generated per prompt and 'the best looking result from each' was selected, due to a high failure rate. This selection procedure should be disclosed in the main text, as it can bias the comparison in favor of DiLightNet; conversely, it makes the comparison less standardized.","section":"Supplemental A.6"},{"comment":"There are several typos and inconsistencies: 'vasts' should be 'vast'; 'generelizability' should be 'generalizability'; 'comaprisons' should be 'comparisons'; 'DilightNet' is inconsistently capitalized; 'morphing' is misspelled in the negative prompt as 'morhping'. Please proofread.","section":"Throughout"},{"comment":"The Limitations section notes that colored light is not supported, global illumination / inter-reflections are not handled, and high-frequency control is moderate. These are significant scope restrictions and should be reflected in the abstract and conclusion, where the claims are currently stated in very general terms.","section":"§6, Limitations"}],"recommendation":"major_revision","confidential_remarks":"The paper has a plausible and interesting core idea, and the efficiency numbers are attractive. However, the evaluation does not support the central 'state-of-the-art control' claim: the control metric is conceptually mismatched, a blending baseline beats the method on that metric, and the user study omits control-specific questions. These are fixable with additional experiments and re-analysis, so I recommend major revision rather than rejection. I would also encourage the authors to make the evaluation dataset and code available immediately, as the self-constructed benchmark is a key part of the contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper before reading it. First, the core idea is clever and the analysis is worth your time: they show that self-attention layers in SD1.5 carry the most light-relevant information, and they use that to train a tiny LoRA regressor on only 4,000 synthetic images to predict direct-irradiance maps, then guide generation with that regressor. The qualitative results across portraits, anime, paintings, and sketches are genuinely diverse, and the efficiency numbers (7.9e5 params vs. 8.6e8 for IC-Light, 4e3 vs. 1e7 data points) are impressive. The ablation study is thorough and the identity-preservation user study is real evidence. Second, the paper's headline claim—state-of-the-art control—is not supported by its own evaluation. The \"Control\" metrics in Tables 1 and 2 compute L2/LPIPS/CLIP/DINO between the relit image and the direct-irradiance map I_c. That is not a picture of the relit scene; it is a shading field on a depth proxy. A correct relit image should not be close to it in pixel or deep-feature space. And the numbers show it: the naive \"Blending\" baseline (50/50 blend of source and I_c) beats the full method on all four control metrics (L2 0.057 vs 0.130), and IC-Light beats the method on every control column. The paper says it is \"on par\" on control adherence; the data says it is worse. The user study never asks about whether the target lighting direction, shadow placement, or highlight shape was achieved—only identity/style and overall relighting. So the control half of the claim rests on a metric that is not measuring control. There are secondary issues: the evaluation dataset is constructed by the authors, control signals come from their own depth-and-render pipeline, hyperparameters are tuned on the same dataset, and no error bars are given. None of these are fatal by themselves, but combined with the metric problem they mean the paper overreaches. That said, the method itself is plausible, the attention analysis is a genuine contribution, and the identity-preservation results look solid. This paper deserves peer review—the idea is worth referee time—but the authors need to redesign the control evaluation or substantially temper the \"state-of-the-art control\" claim. A serious editor should send it out, with the expectation of major revision.","headline":"Good idea and a genuinely interesting analysis of where light lives in SD, but the control-adherence claim is not supported by their own numbers—the control metric is flawed and a naive blend beats the full method on it.","tokens_in":17999,"tokens_out":1895,"would_cite":true,"duration_ms":21293,"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":"Training a tiny LoRA regressor on the self-attention layers of Stable Diffusion 1.5 to predict direct-irradiance maps lets a diffusion model relight arbitrary generated images with roughly 1,000x less data than leading methods.","keywords":["relighting","light control","diffusion models","LoRA","classifier guidance","self-attention","direct irradiance","image editing"],"falsifier":"Render a diverse set of test images with known light positions, run PractiLight's regressor on them, and compare its predicted direct-irradiance map to the ground-truth render; if the prediction's light direction or shadow placement is consistently wrong outside the primitive domain, the generalization claim fails.","tokens_in":17117,"feed_emoji":"💡","tokens_out":4691,"duration_ms":44222,"temperature":0.7,"pith_summary":"The paper tries to establish that the lighting knowledge needed for relighting already lives inside a pretrained text-to-image diffusion model, and that a very small adapter can pull it out. The authors train a lightweight LoRA regressor, placed only on self-attention layers, to output a direct-irradiance map from an image, using 4,000 synthetic renders of simple shapes. At generation time, this regressor acts as a classifier-guidance energy that steers denoising toward a user-specified light condition, while query injection and an edge ControlNet protect identity. If correct, the result is a practical relighting tool that works across portraits, anime, sketches, paintings, and outdoor scenes without domain-specific training data. The paper supports this with quantitative metrics, ablations, and user studies against IC-Light and DiLightNet.","feed_headline":"Self-attention unlocks relighting with 1,000x less data","feed_subtitle":"A small LoRA regressor trained on simple shapes guides light edits across portraits, anime, sketches, and more.","key_machinery":"The direct-irradiance map is the control signal: an image of the first-bounce light arriving at each pixel, combining direct diffuse and direct specular light. The LoRA regressor is a low-rank adapter placed on the self-attention layers of Stable Diffusion 1.5 that predicts such a map from an image, and it is the mechanism that unlocks the foundation model's light knowledge. Classifier guidance uses the regressor's prediction as an energy term during denoising, steering the latent toward the user's desired direct-irradiance. Query injection from the source image's self-attention and an edge-conditioned ControlNet are supporting mechanisms that preserve identity and structure.","core_discovery":"The central discovery is that self-attention activations in a latent diffusion model are the right place to read and write lighting. By measuring how injecting target-light activations into a source image's denoising affects semantic similarity, the authors find self-attention layers, especially decoder ones, carry light transport better than other layers. They also find light-related attention acts early in denoising, then fades around the midpoint. Building on that, they convert the U-Net backbone into an image-to-image regressor via rank-8 LoRA on self-attention q/k/v/output projections, trained to predict clean direct-irradiance maps. Used as classifier guidance with a scheduled scale, t","pith_inferences":["The per-head analysis hints at even finer control: if only some self-attention heads encode light, a per-head LoRA or direct head manipulation could relight with a fraction of the current parameters.","The same attention-based regressor scheme should transfer to transformer-based diffusion models such as Flux, since the paper shows attention propagation in those models follows light bounces; a direct test would be training the same LoRA on a DiT backbone.","Adding colored light, inter-reflections, or multi-bounce rendering to the synthetic training set could remove two stated limitations (no colored light, no global illumination) without changing the method's structure.","The regressor could be reused outside relighting, for example to edit shadows or highlights locally or to serve as a lighting prior for 3D-aware generation, since it outputs a physically meaningful irradiance estimate."],"forward_implications":["Relighting can be performed with three orders of magnitude fewer parameters and data, making it feasible on a single GPU and accessible to non-specialists.","Because the backbone is not fully fine-tuned, the model retains its generalization across image domains rather than collapsing to the training domain.","Lighting control should be applied during early-to-mid diffusion timesteps, giving a practical scheduling rule for future light-editing methods.","Self-attention layers, not cross-attention, are the correct interface for lighting, pointing other editing tasks toward attention-based adapters.","Identity-preserving relighting needs both structural control (ControlNet) and query injection; each alone is insufficient."],"supporting_citations":[{"why":"Supplies the pretrained Stable Diffusion 1.5 latent diffusion backbone whose self-attention layers the method taps.","marker":"[Rombach et al. 2022]"},{"why":"Supplies LoRA, the low-rank adaptation mechanism used to fine-tune the self-attention layers into a regressor.","marker":"[Hu et al. 2021]"},{"why":"Supplies classifier guidance, the mechanism by which the regressor's prediction steers the diffusion sampling process.","marker":"[Dhariwal and Nichol 2021]"},{"why":"Readout Guidance is the prior guidance approach the method adapts and quantitatively compares against.","marker":"[Luo et al. 2024a]"},{"why":"IC-Light is a leading full-model fine-tuning baseline that the paper claims to outperform in quality and efficiency.","marker":"[Zhang et al. 2025]"},{"why":"RGB↔x is the intrinsic-channel decomposition baseline used in quantitative comparisons.","marker":"[Zeng et al. 2024a]"},{"why":"ControlNet, conditioned on edge maps, is used to preserve structure and transfer shadow edges during relighting.","marker":"[Zhang et al. 2023]"},{"why":"Query injection from the source image's self-attention is used to preserve identity and reduce cartoonification.","marker":"[Alaluf et al. 2024]"},{"why":"ZoeDepth provides the monocular depth estimate used to build the pseudo-3D scene for the light-control signal.","marker":"[Bhat et al. 2023]"},{"why":"The GGX microfacet material model is used both in the synthetic training data and in rendering the control irradiance map.","marker":"[Walter et al. 2007]"}],"fun_headline_variants":["Relight any image by tapping diffusion's self-attention","Light control via self-attention: 1,000x less data","Foundational diffusion model controls light with tiny data","Practical relighting: LoRA on self-attention wins","Tiny LoRA, big control: relight via self-attention"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"A regressor trained on only 4,000 synthetic images of a few geometric primitives lit by a single point light can estimate direct-irradiance maps accurately enough for arbitrary generated images—portraits, anime, sketches, paintings—without breaking identity.","fun_headline_variants_meta":{"raw":{"variants":["Relight any image by tapping diffusion's self-attention","Light control via self-attention: 1,000x less data","Foundational diffusion model controls light with tiny data","Practical relighting: LoRA on self-attention wins","Tiny LoRA, big control: relight via self-attention"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000726,"raw_usage":{"total_tokens":3086,"prompt_tokens":738,"completion_tokens":2348,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":482,"completion_tokens_details":{"reasoning_tokens":2263}},"tokens_in":482,"tokens_out":2348,"duration_ms":18105,"temperature":1.0,"reasoning_tokens":2263,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T12:07:42.808840+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Render a diverse set of test images with known light positions, run PractiLight's regressor on them, and compare its predicted direct-irradiance map to the ground-truth render; if the prediction's light direction or shadow placement is consistently wrong outside the primitive domain, the generalization claim fails.","supporting_citations":[],"review_version":1}