{"id":"927eff85-3ddf-42d6-b396-d028cc0f82e5","arxiv_id":"2502.08590","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Light-A-Video relights videos without training by injecting image relight results into a video diffusion model's denoising loop with cross-frame attention and progressive blending.","lead":"This paper presents a method to relight videos without any training, adapting an image relighting model to work across frames with attention sharing and progressive blending. If it works as described, content creators can change lighting in existing footage using off-the-shelf AI models.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's central claim of preserving relighted image quality rests on a self-referential FID: Light-A-Video is compared against the very per-frame IC-Light outputs it uses as guidance (Sec. 5.1), so the metric cannot distinguish quality from failure to smooth.","rationale":"The reader correctly identified the PLF interpolation as a theoretical weakness, and that concern is real: Eq. 10's convex blend is presented as a consequence of light-transport additivity, but the inputs are decoded, non-linear image representations, not linear radiance values, so the physical justification is loose. However, the single most load-bearing concern is the self-referential FID metric, which the reader mentioned in the rationale but did not rank as the weakest assumption. The paper's strongest quantitative evidence for the quality-preservation component of the central claim is Table 1's FID column, and that column compares against the very model whose per-frame outputs the method uses as guidance. A trivial baseline that simply returns IC-Light per-frame results would achieve a near-perfect FID while failing the temporal consistency goal, showing that the metric cannot separate quality from the absence of smoothing. The core method may still work, and the qualitative results and user study are suggestive, but the central claim is not quantitatively supported until the quality metric is made non-circular. Since the reader's verdict was already CONDITIONAL and this concern is consistent with that verdict, no verdict change is needed.","tokens_in":14825,"tokens_out":5104,"duration_ms":58214,"concrete_test":"Replace the reference set in the FID evaluation with real or rendered ground-truth relit frames (e.g., render each test scene under the target HDRI environment). Recompute FID/LPIPS/PSNR for Light-A-Video, IC-Light+SDEdit-0.6, and the trivial per-frame IC-Light baseline. If Light-A-Video's FID/LPIPS to ground truth is not better than the trivial baseline or SDEdit-0.6, the claim of preserving relighted image quality fails. As a minimal sanity check, compute the paper's own FID for the trivial baseline: if it is approximately 0 with a low temporal CLIP score, the current metric is degenerate and cannot support the quality claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 5.1 defines 'Relighted Image Quality' as the FID between each method's frames and the frame-by-frame IC-Light results; Table 1 then reports Light-A-Video's FID of 29.63 as evidence that it maintains quality. This is load-bearing because the central claim has two parts—temporal smoothness and preserved relight quality—and the quality part is supported only by this metric. The reference distribution is the output of the same IC-Light model that the pipeline uses to build the fusion target (Sec. 4.3, Eq. 10). A trivial method that outputs the per-frame IC-Light frames unchanged would score an FID near 0 while being temporally flickery, exactly the failure mode Light-A-Video is designed to remove. Conversely, a method that successfully smooths lighting by altering frames away from IC-Light's independent per-frame outputs is penalized. The CLIP temporal score also rewards temporal self-similarity and does not by itself establish correct lighting; motion preservation only checks motion, not illumination. Thus the quantitative claim that relighting quality is 'maintained' is not established by the reported numbers. A secondary concern is that Eq. 10's interpolation is justified by light-transport linearity (Eqs. 5–6), but I_v and I_r are decoded, non-linear, likely tone-mapped representations, not linear radiance, so the physical derivation does not directly apply. The FID issue, however, is the more decisive gap in the paper's evidence.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Light-A-Video, a training-free pipeline for video relighting that combines a per-frame image relighting model (IC-Light) with a video diffusion model (AnimateDiff or CogVideoX). Two components are introduced: a Consistent Light Attention (CLA) module that stabilizes the generated light source via cross-frame attention, and a Progressive Light Fusion (PLF) strategy that blends the source video appearance with the relighted appearance during denoising under a linear-interpolation schedule. The paper reports improved temporal CLIP scores, lower motion preservation error, and better user-study rankings relative to baselines, and it includes qualitative results on video and foreground-sequence relighting.","tokens_in":15180,"tokens_out":2767,"duration_ms":33001,"significance":"If validated, the method would be a practical, training-free way to extend image relighting models to video, with easy integration into existing video diffusion backbones. The paper is honest about limitations (dynamic lighting changes are not handled) and provides qualitative evidence, a user study, and ablations over the main hyperparameters. However, the central quantitative claim of maintained relighted image quality rests on a self-referential FID metric, and the physical justification for the fusion equation is improperly applied. These issues leave the core contribution plausible but not yet established by the reported experiments.","major_comments":[{"comment":"","section":"Section 5.1, Table 1"},{"comment":"","section":"Section 4.3, Eqs. (5)-(10)"}],"minor_comments":[{"comment":"","section":"Eq. (2)"},{"comment":"","section":"Section 5.1"},{"comment":"","section":"Section 4.2, Eq. (8)"},{"comment":"","section":"Table 1"},{"comment":"","section":"Section 5.6"}],"recommendation":"major_revision","confidential_remarks":"The self-referential FID is the main blocker: it directly undermines the paper's quantitative claim of preserving relighted image quality. The authors should be asked to replace it with a non-circular metric or to provide a strong perceptual study, and to re-examine the physical grounding of Eq. (10). The method itself is interesting and the qualitative results are promising, so the paper is within scope for revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this one. The pipeline is training-free video relighting: run a per-frame image relighter (IC-Light) with a temporal-averaging attention module (CLA), then steer a video diffusion model's denoising trajectory toward a progressively blended relight target (PLF). The combination is new even though the ingredients are not; it works with both AnimateDiff and CogVideoX. The paper is clearly written, the ablation isolates the two modules, and the user study is reasonably sized. The qualitative results look genuinely smoother than frame-by-frame relighting, and I believe the method works.\n\nThe soft spots are real. The FID for \"Relighted Image Quality\" in Table 1 is computed against the per-frame IC-Light outputs. Since IC-Light is inside the pipeline, the metric rewards staying close to IC-Light and penalizes the temporal smoothing the method is supposed to add. A method that simply outputs IC-Light frames unchanged would score near zero FID. So the paper does not establish that relight quality is preserved. The temporal CLIP score measures self-similarity, not lighting correctness, and motion preservation says nothing about illumination. The user study is the remaining evidence for lighting quality; it is favorable, but it is a ranking, not an absolute measure.\n\nSecondary issue: the light-transport justification for Eq. 10 is hand-wavy. Eq. 6 is about adding linear radiance; Eq. 10 is convex interpolation of decoded, non-linear images. The design may work anyway, but the physics as written does not justify the interpolation.\n\nThere are also no error bars, no code, and the hyperparameters (gamma, lambda schedule) are set empirically. These can be fixed in revision.\n\nBottom line: the central idea is plausible and the qualitative results are credible, but the quantitative case is not. I would send this to peer review; the right referee will ask for a non-self-referential quality metric and code release.","headline":"A practical training-free relighting pipeline whose quality claim rests on a self-referential FID; the method itself is plausible and worth refereeing.","tokens_in":15699,"tokens_out":2886,"would_cite":false,"duration_ms":30460,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Light-A-Video makes video relighting training-free by blending frames during denoising.","keywords":["video relighting","training-free","diffusion models","consistent light attention","progressive light fusion","temporal consistency","light transport","IC-Light"],"falsifier":"Use a synthetic scene with known geometry, albedo, and two environment lights, so ground-truth renders exist for every convex combination of the lights. Compare $I_v^t + \\lambda_t(I_r^t - I_v^t)$ against a true render under the blended light for a sweep of $\\lambda_t$; if the frames diverge substantially in shadow or specular regions, the light-transport interpolation behind Eq. 10 is empirically false. A second check is to weaken the VDM's refinement by using a near-zero-strength denoiser and ask whether the fused frames alone form a temporally smooth relighted video.","tokens_in":14639,"feed_emoji":"💡","tokens_out":6765,"duration_ms":69722,"temperature":0.7,"pith_summary":"Light-A-Video sets out to prove that temporally smooth relighting of arbitrary videos can be achieved without any training or optimization, by reusing a frozen image relighting model in tandem with a frozen video diffusion model. The authors show that a straightforward per-frame application of an image relighter flickers because the generated light source and the relighted appearance are unconstrained across time. They counter this with two interventions: consistent light attention, which injects time-averaged features into the relighter's self-attention to stabilize the background light, and progressive light fusion, which replaces the video model's denoising target with a linearly blended version of the source and relighted frames, with the blend weight shrinking as denoising progresses. If the claim holds, relighting becomes an inference-only operation that works with off-the-shelf components, and it supports both whole-video relighting and text-guided relighting of foreground sequences with generated backgrounds.","feed_headline":"Video relighting goes training-free with a progressive light blend","feed_subtitle":"It mixes source and relit frames during denoising so light and motion stay consistent across the clip.","key_machinery":"The argument runs on two modules placed in the inference loop. Consistent Light Attention (CLA) alters each self-attention layer of the frozen IC-Light relighter so that the frame's features are mixed with their temporally averaged counterpart, $h'_o = (1-\\gamma) h'_1 + \\gamma \\bar{h}'_2$, giving the generated light source a stable cross-frame anchor. Progressive Light Fusion (PLF) is the deeper mechanism: for each denoising step the video model's predicted ``consistent target'' $z^v_{0\\leftarrow t}$ is relighted to $z^r_{0\\leftarrow t}$, both are decoded and blended by the light-transport-motivated formula $I_p^t = I_v^t + \\lambda_t(I_r^t - I_v^t)$ with $\\lambda_t = 1 - t/T_m$, and the re-encoded blend $E(I_p^t)$ replaces the original denoising target so that the step direction is bent toward relighting without leaving the video model's refinement envelope. A detail compensation term $\\Delta d_m$ captured in the first denoising step is added back so that source-video details are not lost in the consistent target.","core_discovery":"The central discovery is that the illumination change imposed by an image relighting model can be transferred onto a video simply by steering the denoising trajectory of a video diffusion model; no video-specific relighting network or dataset is required. At each denoising step the video model's predicted clean frame is relit by the image model, and then a convex combination of the original and relighted appearances, $I_p^t = I_v^t + \\lambda_t (I_r^t - I_v^t)$, is re-encoded and used as the new denoising target. The weight $\\lambda_t$ decreases from near 1 to 0 across the schedule, so early steps push the output toward the relighted appearance while later steps let the video model's motion priors snap the result back to a coherent video. The paper presents this as a practical application of the linearity of light transport: since a scene's appearance under combined illumination is the sum of appearances under each illumination, the authors treat appearance blending as a proxy for lighting interpolation, and report that a constant blend target degrades quality while the progressive schedule preserves both relight fidelity and temporal consistency.","pith_inferences":["The same progressive-fusion recipe could generalize to other per-frame image edits—color grading, exposure adjustment, deflickering—by using a different base editor in place of the relighter, since nothing in PLF is lighting-specific beyond the motivation.","The paper's identification of appearance interpolation with lighting interpolation suggests a concrete failure mode the authors do not test: cast shadows and specular highlights move nonlinearly with light direction, so at intermediate $\\lambda_t$ these regions may ghost or blur, and a region-adaptive schedule weighting shadow and specular areas differently could address it.","A testable extension is to modulate $\\lambda_t$ per-pixel or per-region based on albedo: stable regions can be relighted quickly while high-frequency illumination effects are introduced later, which would preserve identity and motion more faithfully.","CLA's temporal average is a fixed blur over the whole clip; replacing it with a temporally shifted or tracked average could address the acknowledged dynamic-lighting limitation and enable relighting under moving light sources."],"forward_implications":["Per-frame image relighting becomes a serviceable video relighter: any user with IC-Light and a public video diffusion model can relight a video or a foreground sequence with a text prompt, with no training run.","The method is backbone-agnostic: the same progressive-fusion recipe operates on UNet-based models such as AnimateDiff and DiT-based models such as CogVideoX, so improvements in either component should transfer directly.","Because the fusion target is a linear blend, the schedule $\\lambda_t = 1 - t/T_m$ is a single hyperparameter; the supplementary experiments show that a constant blend target degrades motion and quality, implying the progressive schedule is what keeps the denoising trajectory inside the video model's refinement capability.","The approach inherits the limitations of its frozen components: output quality tracks the image relighter and the video diffusion model, and the authors note that the CLA module does not handle dynamic lighting changes well."],"supporting_citations":[{"why":"Supplies IC-Light, the frozen image relighting model whose per-frame output is fused by CLA and PLF; the method's relight quality is inherited from it.","marker":"[62]"},{"why":"AnimateDiff is the primary video diffusion backbone, providing motion priors and the denoising trajectory used in the main experiments.","marker":"[13]"},{"why":"CogVideoX is the DiT-based video backbone that demonstrates the approach works beyond UNet architectures.","marker":"[59]"},{"why":"Debevec reflectance-field work grounds the light-transport linearity argument that motivates additive appearance fusion.","marker":"[9]"},{"why":"SDEdit defines the noising-based editing baseline that the method improves on, and its target-replacement idea is the closest prior mechanism.","marker":"[30]"},{"why":"AnyV2V is the appearance-propagation baseline that Light-A-Video compares against and outperforms on motion preservation.","marker":"[25]"},{"why":"RAFT estimates optical flow for the motion-preservation metric used to validate temporal consistency.","marker":"[49]"}],"fun_headline_variants":["Training-free video relighting via progressive light fusion","Relight videos without training using progressive light blend","Progressive light fusion enables training-free video relighting","No training needed: progressive light blend for video relighting","Flicker-free video relighting with progressive light fusion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a pixel-space blend of the source frame and the relighted frame, with a weight that shrinks during denoising, behaves like an intermediate illumination state that the video model can refine; if that blend is not a valid middle state, the denoising trajectory is steered toward an appearance that the model must then conceal rather than complete.","fun_headline_variants_meta":{"raw":{"variants":["Training-free video relighting via progressive light fusion","Relight videos without training using progressive light blend","Progressive light fusion enables training-free video relighting","No training needed: progressive light blend for video relighting","Flicker-free video relighting with progressive light fusion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000283,"raw_usage":{"total_tokens":1707,"prompt_tokens":1017,"completion_tokens":690,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":633,"completion_tokens_details":{"reasoning_tokens":614}},"tokens_in":633,"tokens_out":690,"duration_ms":6915,"temperature":1.0,"reasoning_tokens":614,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T04:32:20.179733+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Use a synthetic scene with known geometry, albedo, and two environment lights, so ground-truth renders exist for every convex combination of the lights. Compare $I_v^t + \\lambda_t(I_r^t - I_v^t)$ against a true render under the blended light for a sweep of $\\lambda_t$; if the frames diverge substantially in shadow or specular regions, the light-transport interpolation behind Eq. 10 is empirically false. A second check is to weaken the VDM's refinement by using a near-zero-strength denoiser and ask whether the fused frames alone form a temporally smooth relighted video.","supporting_citations":[{"cited_title":"Scal- ing in-the-wild training for diffusion-based illumination har- monization and editing by imposing consistent light trans- port","cited_arxiv_id":null,"evidence_quote":"Supplies IC-Light, the frozen image relighting model whose per-frame output is fused by CLA and PLF; the method's relight quality is inherited from it."},{"cited_title":"Acquiring the reflectance field of a human face","cited_arxiv_id":null,"evidence_quote":"Debevec reflectance-field work grounds the light-transport linearity argument that motivates additive appearance fusion."}],"review_version":1}