Pith. sign in

REVIEW 4 major objections 6 minor 70 references

LiveLight is the first diffusion-based video relighting system to run in real time: a user can drag a 3D light and see temporally coherent relit frames stream out at about 16 FPS, on arbitrarily long videos.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review

2026-08-04 21:11 UTC pith:ARFV3ERX

load-bearing objection Real-time streaming relighting is a genuine integration win, but the SOTA claim is undercut by the missing RelightMaster comparison and an ambiguous synthetic benchmark that may use oracle depth. the 4 major comments →

arxiv 2608.01771 v1 pith:ARFV3ERX submitted 2026-08-03 cs.CV

LiveLight: Real-time Streaming Video Relighting with Interactive Control

Yue Ma , Jiangming Wang , Yucheng Wang , Xilai Wang , Zhiyuan Li , Xinyu Wang , Hongyu Liu , Ruofan Liang
show 3 more authors
Songchun Zhang Yuxuan Xue Qifeng Chen
This is my paper
classification cs.CV
keywords video relightingdiffusion modelsreal-time generationstreaming videointeractive lighting controlMulti-Plane Light Irradiancefew-step distillationtemporal coherence
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper sets out to move video relighting from an offline, clip-level render to a live creative instrument: instead of specifying the whole lighting trajectory up front and waiting for a batch result, a user drags a virtual 3D light and watches the relit frames stream out. Its central claim is that a diffusion model can do this at interactive speed—roughly 16 frames per second, with a 16-frame clip processed in about one second of latency—while producing output that is more temporally stable and more controllable than offline baselines, on both short clips and 256- to 512-frame videos. Three mechanisms carry the claim: a lightweight adapter that injects depth-aware Multi-Plane Light Irradiance maps into a pretrained diffusion backbone; a training-only geometry feedback branch that keeps shading consistent with estimated depth and normals despite a four-step denoising budget; and a rolling-window denoising ladder that emits each chunk once while retaining latent context, so latency stays bounded and stream length does not. If the claim is right, relighting stops being a re-render and becomes a real-time parameter you can turn while the video plays.

Core claim

LiveLight shows that real-time interactive video relighting is achievable inside a diffusion model by replacing the offline 'trajectory first, render later' workflow with a streaming one. Three designs work together: a 12-channel depth-aware Multi-Plane Light Irradiance condition, built by projecting the user's light onto four fronto-parallel depth planes and injected through a one-million-parameter cross-attention adapter, keeps light position, intensity, and color explicit and editable per frame; a training-only geometry feedback branch, supervised by a frozen depth-and-normal estimator, preserves geometric plausibility when inference is capped at four denoising steps; and a rolling window

What carries the argument

The central object is the Multi-Plane Light Irradiance (MPLI) condition: the user's light is projected onto four fronto-parallel depth planes around the subject, yielding a 12-channel, depth-aware tensor that encodes 3D light geometry. A ~1M-parameter adapter turns this tensor into lighting tokens, injected by cross-attention into the shallow layers of a Stable Diffusion-based backbone, so updating the light only means recomputing and re-encoding the MPLI map. Two further mechanisms carry speed and streaming: a training-only geometry feedback branch (Eq. 4) back-propagates depth and normal errors from a frozen estimator, keeping four-step denoising geometrically plausible at zero inference c

Load-bearing premise

The entire pipeline depends on a pretrained depth-and-normal estimator: it builds the lighting condition at inference and supervises training-time geometry, so wherever that estimator is wrong—complex hairstyles, occlusions, extreme poses—the claimed relighting fidelity degrades.

What would settle it

Run LiveLight on videos where the monocular depth estimator is known to fail (heavy occlusion, complex hair, unusual poses) and compare against a synthetic ground-truth relighting render: if fidelity (PSNR/SSIM, shadow placement) does not drop relative to simple-geometry control videos, the depth prior is less load-bearing than claimed; if it collapses, the central claim fails. A cleaner quantitative probe: perturb the depth map used to build the MPLI condition (e.g., ±10% plane shifts) and measure the drop in lighting-tracking accuracy and temporal consistency—that drop measures how much of t

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Video relighting becomes a live parameter: position, intensity, and color of a light can be changed while the output is streaming, with the model following per frame instead of per clip.
  • Temporal coherence over long videos is achieved without overlap re-generation or blending, so latency stays roughly constant (~1.01 s per 16-frame clip) while stream length is unbounded.
  • The TinyVAE variant raises throughput to 18.34 FPS with marginal quality loss, showing the pipeline can trade decoder fidelity for interactivity whenever latency is the priority.
  • Geometry-guided feedback offers a recipe for few-step diffusion distillation that does not sacrifice shading plausibility, because the constraints are imposed offline and add no inference cost.
  • Releasing the models, training data, and the synthetic UE5 generator would give the community a shared, controllable benchmark for interactive relighting.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Because the lighting condition is recomputed per frame, LiveLight could double as an interactive light-scripting tool—animating a light along a keyframed path or syncing color shifts to audio—which the paper only gestures at as 'dynamic control'.
  • The admitted failure modes (complex hairstyles, occlusions, extreme poses) all trace to the depth prior, so coupling the pipeline to a more robust or user-correctable geometry source is the most direct route to broader coverage; the paper leaves that coupling unexplored.
  • None of the three designs except the MPLI map itself is lighting-specific, so the same recipe—compact physical condition, training-only geometry feedback, rolling-window ladder—plausibly transfers to other per-frame controllable video edits such as camera or pose.
  • A stress test the paper does not report: measuring appearance and lighting drift over streams of thousands of frames, to quantify how tightly the rolling window bounds error accumulation relative to overlap-blending alternatives.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper presents LiveLight, a diffusion-based framework for real-time streaming video relighting with interactive 3D lighting control. The method has three components: (i) a lightweight adapter that injects a Multi-Plane Light Irradiance (MPLI) condition into the diffusion backbone; (ii) few-step diffusion distillation with a geometry-guided feedback branch that supervises depth and normal consistency at training time; and (iii) a rolling-window streaming strategy over latent micro-chunks that emits frames with per-frame reference refresh. The model is trained on a large UE5-rendered paired dataset. Experiments compare LiveLight with AnyV2V, LightCtrl, Light-A-Video, and TC-Light on real Internet videos, synthetic UE5 sequences, and a user study, reporting 15.78 FPS with the standard SD VAE (18.34 FPS with TinyVAE) and 1.01 s per 16-frame clip. Section 5 candidly lists limitations, including reliance on off-the-shelf depth/normal estimators and artifacts on small faces and high-frequency details.

Significance. If the reported results hold, LiveLight would be a meaningful step toward interactive, streaming video relighting: the rolling-window latent propagation is a sensible way to bound latency while maintaining temporal coherence, and the training-time geometry feedback is a clean way to preserve structure without adding inference cost. The three-stage ablation study is internally consistent, and the authors' promise to release models, data, and the synthetic generator would be a valuable community contribution. However, the evaluation as written does not yet support the headline 'state-of-the-art' and 'significantly outperforming' claims: the closest baseline is omitted, the synthetic evaluation may use oracle geometry, the quantitative metrics lack statistical grounding, and real-world transfer is not quantitatively tested.

major comments (4)
  1. [§3.2, Eq. (2); §4.3, Tables 1–3] The MPLI condition is taken from RelightMaster [1], which is itself a video relighting method, yet it appears in references only as the representation source and is never included in any quantitative comparison. Without this baseline, the claim that LiveLight achieves 'state-of-the-art relighting quality' is not testable. Please add RelightMaster to Tables 1–3, or explain why it is not a valid baseline for this setting.
  2. [§3.2, Eq. (2); §4.3, Table 2; §5] The synthetic margin in Table 2 (43.264 PSNR vs 39.726 for TC-Light) is only meaningful if the deployed geometry pipeline is used at evaluation. The text does not state whether the MPLI condition in Table 2 is built from the off-the-shelf monocular depth estimator or from the UE5-exported ground-truth depth. Training-data synthesis (Sec. 3.1) exports GT depth, while Eq. (2) refers to 'estimated subject depth d_i'; Sec. 5 admits the estimator is unreliable for complex hair, occlusions, and extreme poses. If GT depth was used, the result measures an oracle-geometry system. Please specify the evaluation protocol and, if GT depth was used, rerun with the deployed estimator and report the drop.
  3. [§4.3, Tables 1–3; user study] The quantitative claims are single numbers with no confidence intervals or significance tests; several VBench margins are small (e.g., Temporal Cons. 0.9872 vs 0.9799 in Table 1), so 'significantly outperforming' is not established. The two relighting-specific metrics in Table 1 are nonstandard: PSNR_y is computed against a pure-white reference inside a mask, which rewards over-brightening rather than correct illumination, and PSNR_light is computed against an overlay of the target light map, not against calibrated ground truth. The synthetic test set (Table 2) is rendered by the same UE5 pipeline used for training, so it is not an independent test of real-world transfer. Add error bars/paired tests, validate the custom metrics, and evaluate on real-world data with known lighting.
  4. [§3.4; §4.3, Table 1] Real-time interactivity is the central claim, but only FPS and total clip time are reported. With a rolling window of K micro-chunks, chunk size M=4, and per-frame reference refresh, end-to-end latency from a user light change to the displayed output is not determined by throughput alone. Please report startup latency, per-frame incremental latency, and memory usage for arbitrarily long streams.
minor comments (6)
  1. [§3.4] The text says 'Details are provided in Appendix??' — this unresolved reference should be fixed before submission.
  2. [Tables 1–3; Fig. 6] The baseline 'AnyV2V' is cited as [64], but reference [64] is the IC-Light paper. Provide the correct AnyV2V citation and clarify the IC-Light+AnyV2V pipeline described in Sec. 4.3.
  3. [Fig. 3b; §3.3] Fig. 3b shows a 'Stop Grad' path, but the text does not specify exactly which branches are detached during the m-step distillation rollout. Clarify the gradient flow.
  4. [Eq. (6)] The Stage 1 objective uses \hat{z}_0 and z_0 without defining the prediction head or whether this is a z0-prediction variant of the standard SD loss. Make the objective and the role of \tau and \epsilon precise.
  5. [Table 2; Fig. 5] Table 2 has the typo '10 different scene' (should be 'scenes'). Fig. 5 reports relative improvements but the axes and raw curves are not labeled.
  6. [§3.2; §5] The specific off-the-shelf depth/normal estimator used for MPLI construction and geometry feedback is never named. Specify the architecture and source of this estimator, since Sec. 5 identifies it as a source of failure.

Circularity Check

0 steps flagged

No circularity: the derivation is a forward-trained conditional pipeline; evaluation and support concerns are not definitional circularity.

full rationale

LiveLight's derivation chain is not circular. The MPLI condition (Eq. 2) is an input representation adopted from RelightMaster [1], not a quantity the paper claims to derive; the lightweight adapter (Eq. 3), few-step distillation (Sec. 3.3), geometry-guided feedback (Eq. 4), and rolling-window scheme (Eq. 5) are all trained end-to-end on paired synthetic renders and then evaluated forward. No parameter is fitted to the evaluation set, and no reported 'prediction' reduces by construction to a training input or to a fitted value. The synthetic benchmark in Table 2 comes from the same UE5 generator as the training data, but the paper states the evaluation scenes and lights are unseen; this is a distribution-match and benchmark-fairness concern, not a definitional circularity. The more salient issues—omission of RelightMaster [1] from comparisons, ambiguity about whether synthetic evaluation uses ground-truth depth or the deployed monocular estimator, the unresolved 'Appendix??' reference in Sec. 3.4, and the acknowledged geometry-estimator limitation in Sec. 5—are correctness and evidence concerns, not circular steps. Self-citations appear in related work and background but are not load-bearing; the load-bearing MPLI citation [1] is not authored by the present paper's authors. Therefore the central claim retains independent content.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The central claim rests on a large synthetic dataset and an off-the-shelf geometry estimator; the paper contributes a system integration, not new physics or entities. The free parameters are mostly standard hyperparameters, with the most consequential choices (P, M, denoising steps) guided by ablations.

free parameters (6)
  • Number of MPLI planes P = 4
    Set to 4 based on ablation across P=2,4,8 (Table 5). The choice affects depth-dependent lighting controllability and model capacity.
  • Micro-chunk size M = 4
    Set to 4 based on ablation across M=2,4,8 (Table 5). It trades temporal coherence against latency and memory.
  • Number of distillation denoising steps = 4
    Set to 4 based on ablation across 2,4,8 steps (Table 5). More steps improve quality marginally but reduce FPS; fewer steps introduce artifacts.
  • Reconstruction and geometry loss weights (lambda_mse, lambda_lpips, lambda_geo, lambda_N) = not reported
    Weights are used in Eqs. (7) and (8) but their numerical values are not given in the paper. They are presumably tuned by hand.
  • Light intensity range for synthetic data = [0.5, 1.5] normalized UE units
    Chosen in Section 3.1 to sample target lighting conditions; affects the distribution of relighting effects seen in training.
  • Color vs white illumination ratio in synthetic data = 6:4
    Chosen in Section 3.1 to balance colored and white light training examples.
axioms (5)
  • domain assumption MPLI representation is a sufficient carrier of lighting information for video relighting
    The entire method builds on the MPLI condition from RelightMaster [1] (Section 3.2, Eq. 2). The paper does not justify why MPLI captures the relevant physical lighting effects.
  • domain assumption Pretrained Stable Diffusion and AnimateDiff weights provide useful priors that survive fine-tuning
    Stage 1 initializes from Stable Diffusion and Stage 3 from AnimateDiff motion modules (Section 4.2). The success depends on these priors being preserved during relighting adaptation.
  • domain assumption The frozen off-the-shelf geometry estimator provides reliable depth and normals on synthetic and real images
    The geometry-guided feedback branch (Section 3.3, Eq. 4) and the MPLI construction (Eq. 2) rely on depth/normal estimates. The paper's own Limitation section admits this can fail on complex hairstyles, occlusions, and extreme poses.
  • domain assumption Synthetic UE5 renders with domain randomization transfer to real-world videos
    All paired training data comes from a self-built UE5 renderer (Section 3.1). There is no quantitative real-world ground truth; the paper shows only qualitative real-video results, so the domain transfer is assumed.
  • standard math Standard diffusion denoising and distillation equations
    The training objectives in Section 3.5 use standard L2 denoising and perceptual losses. These are accepted background math.

reviewed 2026-08-04 · how reviews work

0 comments
read the original abstract

We present LiveLight, the first diffusion-based framework for real-time streaming video relighting with interactive 3D lighting control. Achieving this is non-trivial, as it requires overcoming three critical challenges: effectively injecting dynamic 3D lighting into a diffusion model, maintaining high-fidelity generation under an extremely low NFE (Number of Function Evaluations) budget for real-time speed, and facilitating continuous streaming for interactive control. To address these pain points, we propose three key designs. First, for accurate lighting injection, we propose a lightweight adapter that feeds Multi-Plane Light Irradiance (MPLI) conditions-depth-aware irradiance maps encoding 3D lighting geometry-directly into the diffusion backbone. Second, to prevent rendering quality degradation at low NFEs towards real-time distillation, we introduce a geometry-guided feedback branch. This training-time constraint leverages a frozen geometry estimator to enforce depth- and normal-consistent relighting, ensuring geometrically plausible shading without adding inference overhead. Finally, to enable streaming interaction, we develop a progressive rolling-window strategy that maintains a denoising ladder of latent chunks at varying noise levels. By propagating intermediate states, this strategy guarantees temporal coherence and supports arbitrarily long video relighting with per-frame reference refresh. Extensive experiments on real-world and synthetic benchmarks demonstrate that LiveLight achieves state-of-the-art relighting quality while running at real-time speed, significantly outperforming offline baselines in temporal stability, lighting controllability, and user preference. To foster real-time interactive relighting research, we will publicly release our models, training data, and synthetic data generator.

Figures

Figures reproduced from arXiv: 2608.01771 by Hongyu Liu, Jiangming Wang, Qifeng Chen, Ruofan Liang, Songchun Zhang, Xilai Wang, Xinyu Wang, Yucheng Wang, Yue Ma, Yuxuan Xue, Zhiyuan Li.

Figure 1
Figure 1. Figure 1: Showcase of the proposed LiveLight. LiveLight enables real-time streaming video relighting with interactive 3D point-light control, supporting dynamic adjustment of light position, intensity, and color while preserving temporal coherence and visual quality. Abstract We present LiveLight, the first diffusion-based framework for real-time streaming video relighting with interactive 3D lighting control. Achie… view at source ↗
Figure 2
Figure 2. Figure 2: Motivation for LiveLight. Previous offline video re￾lighting methods require the entire input video and a pre-defined lighting trajectory before generation, producing a fixed-length out￾put after a long waiting time. In contrast, LiveLight processes in￾put frames in a streaming fashion with per-frame dynamic lighting control, producing relit frames with low latency and supporting ar￾bitrarily long video ge… view at source ↗
Figure 3
Figure 3. Figure 3: Overview of LiveLight. (a) We first perform lightweight multi-plane image relighting by constructing a multi-plane irradiance representation from the target light and estimated depth. The resulting MPLI condition is encoded by the light encoder and injected into the diffusion relighting backbone together with the reference image. (b) To improve structure-preserving relighting, we introduce a geometry￾guide… view at source ↗
Figure 4
Figure 4. Figure 4: Overview of our synthetic relighting data synthe￾sis pipeline. We randomly combine diverse 3D assets (humans, animals, and objects) with animations and indoor/outdoor scenes, and render them in Unreal Engine 5 under controllable point lights with varying positions, intensities, and colors. For each frame, the renderer exports the RGB image along with depth map and surface normal map for geometry-aware supe… view at source ↗
Figure 5
Figure 5. Figure 5: Geometry-guided feedback for structure-preserving relighting. Compared with the baseline without geometry-guided feedback, our model maintains lower LPIPS and MSE, especially in the later training stage. Over the last 10% of steps, these met￾rics are 7.8% and 19.0% better, respectively. This indicates that depth-normal feedback provides an effective geometric regulariza￾tion toward consistent relighting. t… view at source ↗
Figure 6
Figure 6. Figure 6: Comparison of previous lighting control methods. LiveLight outperforms existing methods in terms of controllability, temporal stability, and real-time efficiency. geometry-guided feedback branch and fine-tune the relight￾ing model with geometry-aware supervision. This stage is trained for 20K iterations with a batch size of 16, keeping the geometry estimator frozen. In Stage 3, we insert tempo￾ral attentio… view at source ↗
Figure 7
Figure 7. Figure 7: Visual results of real-time lighting control with LiveLight. LiveLight enables high-quality relighting video generation under target lighting while preserving the content and motion of the input video [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗
Figure 8
Figure 8. Figure 8 [PITH_FULL_IMAGE:figures/full_fig_p009_8.png] view at source ↗
Figure 10
Figure 10. Figure 10: User study. We conduct a user study to evaluate the perceptual quality of relighting videos generated by LiveLight and baseline methods. Participants are asked to rate each result in terms of light control, appearance consistency, temporal consistency, and motion fidelity. LiveLight receives the highest preference across all criteria, demonstrating its ability to produce controllable, visually consistent,… view at source ↗
Figure 11
Figure 11. Figure 11: Visual results of diverse color lighting control with LiveLight. LiveLight supports real-time lighting control with different colors and positions, producing temporally coherent relighting results across dynamic videos. that the VAE decoder is not the quality bottleneck and can be swapped for faster alternatives when latency is priori￾tized. In all other experiments, we use the standard SD VAE for fair co… view at source ↗
Figure 13
Figure 13. Figure 13: Visual ablation study of the rolling-window strategy. Removing the rolling-window strategy leads to temporal flickering and unstable lighting across chunks. By propagating through the rolling window, our method maintains coherent illumination and stable appearance over time. Input Num=2 Num=4 Num=8 [PITH_FULL_IMAGE:figures/full_fig_p012_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Visual ablation study of the number of MPLI planes. With P=2, the model cannot capture depth-dependent lighting, limiting 3D controllability. P=8 yields marginal gains over P=4 with added parameters. We adopt P=4 as the default. the rolling-window inference process during training. As shown in [PITH_FULL_IMAGE:figures/full_fig_p012_14.png] view at source ↗
Figure 17
Figure 17. Figure 17: Visual ablation study of distillation denoising steps. Distilling to 2 steps (23.17 fps) introduces visible facial artifacts. Using 8 steps (9.43 fps) yields marginal quality gains over 4 steps (15.78 fps). We adopt 4 steps as the default for optimal quality– speed balance. Input Channel concat Ours ControlNet 18.21fps 15.78fps 10.13fps [PITH_FULL_IMAGE:figures/full_fig_p013_17.png] view at source ↗
Figure 15
Figure 15. Figure 15: Visual ablation study of micro-chunk size. M=2 (16.61 fps) causes severe temporal instability. M=8 (11.17 fps) improves coherence but reduces throughput. We adopt M=4 (15.78 fps) for the best latency–quality balance. Input SD VAE Tiny VAE 15.78fps 18.34fps [PITH_FULL_IMAGE:figures/full_fig_p013_15.png] view at source ↗
Figure 19
Figure 19. Figure 19: Visual ablation study of the geometry-guided feed￾back branch. Removing depth (W/o Depth) or normal (W/o Normal) supervision produces highlights and shadows inconsis￾tent with the surface geometry. The full feedback yields the most geometrically plausible relighting. Input Result Input Result [PITH_FULL_IMAGE:figures/full_fig_p013_19.png] view at source ↗
Figure 16
Figure 16. Figure 16: Visual ablation study of VAE decoder. The Tiny VAE (18.34 fps) produces visual quality comparable to the SD VAE (15.78 fps) while offering faster decoding, making it well suited for real-time streaming. sue. Finally, the geometry-guided feedback branch relies on the accuracy of off-the-shelf depth and normal estimators. When the geometry estimation is inaccurate—particularly for complex hairstyles, occlus… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

70 extracted references · 36 canonical work pages · 3 internal anchors

  1. [1]

    Relightmaster: Precise video relighting with multi-plane light images.arXiv preprint arXiv:2511.06271,

    Weikang Bian, Xiaoyu Shi, Zhaoyang Huang, Jianhong Bai, Qinghe Wang, Xintao Wang, Pengfei Wan, Kun Gai, and Hongsheng Li. Relightmaster: Precise video relighting with multi-plane light images.arXiv preprint arXiv:2511.06271,

    arXiv

  2. [2]

    Stable video diffusion: Scaling latent video diffusion models to large datasets.arXiv preprint arXiv:2311.15127, 2023

    Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram V oleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets.arXiv preprint arXiv:2311.15127, 2023. 3

    Pith/arXiv arXiv 2023

  3. [3]

    Real-time 3d-aware portrait video relighting

    Ziqi Cai, Kaiwen Jiang, Shu-Yu Chen, Yu-Kun Lai, Hongbo Fu, Boxin Shi, and Lin Gao. Real-time 3d-aware portrait video relighting. InProceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition, pages 6221–6231, 2024. 4

    work page 2024

  4. [4]

    SynthLight: Portrait Relighting with Diffusion Model by Learning to Re-render Synthetic Faces

    Sumit Chaturvedi, Mengwei Ren, Yannick Hold-Geoffroy, Jingyuan Liu, Julie Dorsey, and Zhixin Shu. Synthlight: Por- trait relighting with diffusion model by learning to re-render synthetic faces.arXiv preprint arXiv:2501.09756, 2025. 3

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  5. [5]

    Diffusion forcing: Next-token prediction meets full-sequence diffu- sion.Advances in Neural Information Processing Systems, 37:24081–24125, 2024

    Boyuan Chen, Diego Mart ´ı Mons ´o, Yilun Du, Max Sim- chowitz, Russ Tedrake, and Vincent Sitzmann. Diffusion forcing: Next-token prediction meets full-sequence diffu- sion.Advances in Neural Information Processing Systems, 37:24081–24125, 2024. 3

    work page 2024

  6. [6]

    Con- textflow: Training-free video object editing via adaptive con- text enrichment.arXiv preprint arXiv:2509.17818, 2025

    Yiyang Chen, Xuanhua He, Xiujun Ma, and Yue Ma. Con- textflow: Training-free video object editing via adaptive con- text enrichment.arXiv preprint arXiv:2509.17818, 2025. 3

    arXiv 2025

  7. [7]

    Rendering synthetic objects into real scenes: Bridging traditional and image-based graphics with global illumination and high dynamic range photography

    Paul Debevec. Rendering synthetic objects into real scenes: Bridging traditional and image-based graphics with global illumination and high dynamic range photography. InAcm siggraph 2008 classes, pages 1–10. 2008. 3

    work page 2008

  8. [8]

    Acquiring the reflectance field of a human face

    Paul Debevec, Tim Hawkins, Chris Tchou, Haarm-Pieter Duiker, Westley Sarokin, and Mark Sagar. Acquiring the reflectance field of a human face. InProceedings of the 27th annual conference on Computer graphics and interac- tive techniques, pages 145–156, 2000. 3

    work page 2000

  9. [9]

    Scaling recti- fied flow transformers for high-resolution image synthesis

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas M ¨uller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling recti- fied flow transformers for high-resolution image synthesis. InForty-first international conference on machine learning,

  10. [10]

    Relightvid: Temporal-consistent diffusion model for video relighting.arXiv preprint arXiv:2501.16330, 2025

    Ye Fang, Zeyi Sun, Shangzhan Zhang, Tong Wu, Yinghao Xu, Pan Zhang, Jiaqi Wang, Gordon Wetzstein, and Dahua Lin. Relightvid: Temporal-consistent diffusion model for video relighting.arXiv preprint arXiv:2501.16330, 2025. 2, 3

    Pith/arXiv arXiv 2025

  11. [11]

    Dit4edit: Dif- fusion transformer for image editing

    Kunyu Feng, Yue Ma, Bingyuan Wang, Chenyang Qi, Haozhe Chen, Qifeng Chen, and Zeyu Wang. Dit4edit: Dif- fusion transformer for image editing. InProceedings of the AAAI Conference on Artificial Intelligence, pages 2969– 2977, 2025. 3

    2025

  12. [12]

    One step diffusion via shortcut models.arXiv preprint arXiv:2410.12557, 2024

    Kevin Frans, Danijar Hafner, Sergey Levine, and Pieter Abbeel. One step diffusion via shortcut models.arXiv preprint arXiv:2410.12557, 2024. 3

    Pith/arXiv arXiv 2024

  13. [13]

    Pai-studio: Cinematic video background replacement with camera-aware motion

    Heyuan Gao, Bangxun Tang, Yiren Song, Guian Fang, Zijian He, Jie Yang, and Mike Zheng Shou. Pai-studio: Cinematic video background replacement with camera-aware motion. arXiv preprint arXiv:2606.01399, 2026. 4

    Pith/arXiv arXiv 2026

  14. [14]

    Animatediff: Animate your personalized text- to-image diffusion models without specific tuning.arXiv preprint arXiv:2307.04725, 2023

    Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text- to-image diffusion models without specific tuning.arXiv preprint arXiv:2307.04725, 2023. 8

    Pith/arXiv arXiv 2023

  15. [15]

    Cameractrl: Enabling camera control for text-to-video generation.arXiv preprint arXiv:2404.02101, 2024

    Hao He, Yinghao Xu, Yuwei Guo, Gordon Wetzstein, Bo Dai, Hongsheng Li, and Ceyuan Yang. Cameractrl: Enabling camera control for text-to-video generation.arXiv preprint arXiv:2404.02101, 2024. 4

    Pith/arXiv arXiv 2024

  16. [16]

    Unirelight: Learning joint decomposition and synthesis for video relight- ing.Advances in Neural Information Processing Systems, 38:108435–108459, 2026

    Kai He, Ruofan Liang, Jacob Munkberg, Jon Hasselgren, Nandita Vijaykumar, Alexander Keller, Sanja Fidler, Igor Gilitschenski, Zan Gojcic, and Zian Wang. Unirelight: Learning joint decomposition and synthesis for video relight- ing.Advances in Neural Information Processing Systems, 38:108435–108459, 2026. 3

    work page 2026

  17. [17]

    Self forcing: Bridging the train-test gap in autoregressive video diffusion.Advances in Neural Informa- tion Processing Systems, 38:167283–167308, 2026

    Xun Huang, Zhengqi Li, Guande He, Mingyuan Zhou, and Eli Shechtman. Self forcing: Bridging the train-test gap in autoregressive video diffusion.Advances in Neural Informa- tion Processing Systems, 38:167283–167308, 2026. 3

    work page 2026

  18. [18]

    Vbench: Comprehensive bench- mark suite for video generative models

    Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive bench- mark suite for video generative models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807–21818, 2024. 9, 10, 11

    work page 2024

  19. [19]

    Neural gaffer: Relighting any object via diffusion.Advances in Neu- ral Information Processing Systems, 37:141129–141152,

    Haian Jin, Yuan Li, Fujun Luan, Yuanbo Xiangli, Sai Bi, Kai Zhang, Zexiang Xu, Jin Sun, and Noah Snavely. Neural gaffer: Relighting any object via diffusion.Advances in Neu- ral Information Processing Systems, 37:141129–141152,

  20. [20]

    Switchlight: Co-design of physics- driven architecture and pre-training framework for human portrait relighting

    Hoon Kim, Minje Jang, Wonjun Yoon, Jisoo Lee, Donghyun Na, and Sanghyun Woo. Switchlight: Co-design of physics- driven architecture and pre-training framework for human portrait relighting. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 25096–25106, 2024

    work page 2024

  21. [21]

    In- trinsic image diffusion for indoor single-view material es- timation

    Peter Kocsis, Vincent Sitzmann, and Matthias Nießner. In- trinsic image diffusion for indoor single-view material es- timation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5198– 5208, 2024. 3

    work page 2024

  22. [22]

    Streamdiffusion: A pipeline-level solution for real-time in- teractive generation

    Akio Kodaira, Chenfeng Xu, Toshiki Hazama, Takanori Yoshimoto, Kohei Ohno, Shogo Mitsuhori, Soichi Sugano, Hanying Cho, Zhijian Liu, Masayoshi Tomizuka, et al. Streamdiffusion: A pipeline-level solution for real-time in- teractive generation. InProceedings of the IEEE/CVF In- ternational Conference on Computer Vision, pages 12371– 12380, 2025. 3

    work page 2025

  23. [23]

    Hunyuanvideo: A systematic framework for large video generative models.arXiv preprint arXiv:2412.03603, 2024

    Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, et al. Hunyuanvideo: A systematic framework for large video generative models.arXiv preprint arXiv:2412.03603, 2024. 3

    Pith/arXiv arXiv 2024

  24. [24]

    Looking backward: Streaming video-to-video translation with feature banks

    Feng Liang, Akio Kodaira, Chenfeng Xu, Masayoshi Tomizuka, Kurt Keutzer, and Diana Marculescu. Looking backward: Streaming video-to-video translation with feature banks. InInternational Conference on Learning Represen- tations, pages 46425–46445, 2025. 3

    work page 2025

  25. [25]

    Diffusion renderer: Neural inverse and forward rendering with video diffusion models

    Ruofan Liang, Zan Gojcic, Huan Ling, Jacob Munkberg, Jon Hasselgren, Chih-Hao Lin, Jun Gao, Alexander Keller, Nandita Vijaykumar, Sanja Fidler, et al. Diffusion renderer: Neural inverse and forward rendering with video diffusion models. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 26069–26080, 2025. 3

    work page 2025

  26. [26]

    Illumicraft: Unified geometry and illumination diffusion for controllable video generation

    Yuanze Lin, Yi-Wen Chen, Yi-Hsuan Tsai, Ronald Clark, and Ming-Hsuan Yang. Illumicraft: Unified geometry and illumination diffusion for controllable video generation. Advances in Neural Information Processing Systems, 38: 27798–27829, 2026. 3

    work page 2026

  27. [27]

    OPSD-V: On-Policy Self-Distillation for Post-Training Few-Step Autoregressive Video Generators

    Hongyu Liu, Chun Wang, Feng Gao, Xuanhua He, Yue Ma, Ziyu Wan, Yong Zhang, Xiaoming Wei, and Qifeng Chen. Opsd-v: On-policy self-distillation for post-training few-step autoregressive video generators.arXiv preprint arXiv:2607.08766, 2026. 3

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  28. [28]

    Flow straight and fast: Learning to generate and transfer data with rectified flow.arXiv preprint arXiv:2209.03003, 2022

    Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow.arXiv preprint arXiv:2209.03003, 2022. 3

    Pith/arXiv arXiv 2022

  29. [29]

    Tc-light: Temporally coherent generative rendering for realistic world transfer.Advances in Neural Information Processing Systems, 38:51463–51489, 2026

    Yang Liu, Chuanchen Luo, Zimo Tang, Yingyan Li, Yuany- ong Ning, Lue Fan, Junran Peng, ZHAO-XIANG ZHANG, et al. Tc-light: Temporally coherent generative rendering for realistic world transfer.Advances in Neural Information Processing Systems, 38:51463–51489, 2026. 2, 8, 9, 10, 11

    work page 2026

  30. [30]

    Follow-your-shape: Shape-aware image edit- ing via trajectory-guided region control.arXiv preprint arXiv:2508.08134, 2025

    Zeqian Long, Mingzhe Zheng, Kunyu Feng, Xinhua Zhang, Hongyu Liu, Harry Yang, Linfeng Zhang, Qifeng Chen, and Yue Ma. Follow-your-shape: Shape-aware image edit- ing via trajectory-guided region control.arXiv preprint arXiv:2508.08134, 2025. 3

    arXiv 2025

  31. [31]

    Follow your pose: Pose- guided text-to-video generation using pose-free videos

    Yue Ma, Yingqing He, Xiaodong Cun, Xintao Wang, Siran Chen, Xiu Li, and Qifeng Chen. Follow your pose: Pose- guided text-to-video generation using pose-free videos. In Proceedings of the AAAI Conference on Artificial Intelli- gence, pages 4117–4125, 2024. 4

    2024

  32. [32]

    Follow-your-emoji: Fine-controllable and expressive freestyle portrait animation

    Yue Ma, Hongyu Liu, Hongfa Wang, Heng Pan, Yingqing He, Junkun Yuan, Ailing Zeng, Chengfei Cai, Heung-Yeung Shum, Wei Liu, et al. Follow-your-emoji: Fine-controllable and expressive freestyle portrait animation. InSIGGRAPH Asia 2024 Conference Papers, pages 1–12, 2024. 3

    2024

  33. [33]

    Controllable video generation: A survey.arXiv preprint arXiv:2507.16869,

    Yue Ma, Kunyu Feng, Zhongyuan Hu, Xinyu Wang, Yucheng Wang, Mingzhe Zheng, Xuanhua He, Chenyang Zhu, Hongyu Liu, Yingqing He, et al. Controllable video generation: A survey.arXiv preprint arXiv:2507.16869,

    arXiv

  34. [34]

    Follow-your-creation: Empowering 4d creation through video inpainting.arXiv preprint arXiv:2506.04590, 2025

    Yue Ma, Kunyu Feng, Xinhua Zhang, Hongyu Liu, David Junhao Zhang, Jinbo Xing, Yinhan Zhang, Ayden Yang, Zeyu Wang, and Qifeng Chen. Follow-your-creation: Empowering 4d creation through video inpainting.arXiv preprint arXiv:2506.04590, 2025. 3

    Pith/arXiv arXiv 2025

  35. [35]

    Follow-your-motion: Video motion transfer via efficient spatial-temporal decoupled finetuning.arXiv preprint arXiv:2506.05207, 2025

    Yue Ma, Yulong Liu, Qiyuan Zhu, Ayden Yang, Kunyu Feng, Xinhua Zhang, Zhifeng Li, Sirui Han, Chenyang Qi, and Qifeng Chen. Follow-your-motion: Video motion transfer via efficient spatial-temporal decoupled finetuning.arXiv preprint arXiv:2506.05207, 2025. 3

    arXiv 2025

  36. [36]

    Follow-your-emoji-faster: To- wards efficient, fine-controllable, and expressive freestyle portrait animation.arXiv preprint arXiv:2509.16630, 2025

    Yue Ma, Zexuan Yan, Hongyu Liu, Hongfa Wang, Heng Pan, Yingqing He, Junkun Yuan, Ailing Zeng, Chengfei Cai, Heung-Yeung Shum, et al. Follow-your-emoji-faster: To- wards efficient, fine-controllable, and expressive freestyle portrait animation.arXiv preprint arXiv:2509.16630, 2025. 4

    arXiv 2025

  37. [37]

    Fastvmt: Eliminat- ing redundancy in video motion transfer.arXiv preprint arXiv:2602.05551, 2026

    Yue Ma, Zhikai Wang, Tianhao Ren, Mingzhe Zheng, Hongyu Liu, Jiayi Guo, Kunyu Feng, Yuxuan Xue, Zix- iang Zhao, Konrad Schindler, et al. Fastvmt: Eliminat- ing redundancy in video motion transfer.arXiv preprint arXiv:2602.05551, 2026. 3

    arXiv 2026

  38. [38]

    Taesd.https://huggingface.co/ madebyollin/taesd, 2023

    madebyollin. Taesd.https://huggingface.co/ madebyollin/taesd, 2023. Hugging Face model repos- itory. 10

    work page 2023

  39. [39]

    Lux post facto: Learning portrait performance relighting with conditional video diffusion and a hybrid dataset

    Yiqun Mei, Mingming He, Li Ma, Julien Philip, Wenqi Xian, David M George, Xueming Yu, Gabriel Dedic, Ahmet Lev- ent Tas ¸el, Ning Yu, et al. Lux post facto: Learning portrait performance relighting with conditional video diffusion and a hybrid dataset. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 5510–5522, 2025. 4

    work page 2025

  40. [40]

    Total relighting: learning to relight portraits for background replacement.ACM Trans

    Rohit Pandey, Sergio Orts-Escolano, Chloe Legendre, Chris- tian Haene, Sofien Bouaziz, Christoph Rhemann, Paul E De- bevec, and Sean Ryan Fanello. Total relighting: learning to relight portraits for background replacement.ACM Trans. Graph., 40(4):43–1, 2021. 3

    2021

  41. [41]

    Lightctrl: Training-free controllable video relighting.arXiv preprint arXiv:2603.27083, 2026

    Yizuo Peng, Xuelin Chen, Kai Zhang, and Xiaodong Cun. Lightctrl: Training-free controllable video relighting.arXiv preprint arXiv:2603.27083, 2026. 2, 3, 8, 9, 10, 11

    work page arXiv 2026

  42. [42]

    Relightful harmonization: Lighting-aware portrait background replacement

    Mengwei Ren, Wei Xiong, Jae Shin Yoon, Zhixin Shu, Jianming Zhang, HyunJoon Jung, Guido Gerig, and He Zhang. Relightful harmonization: Lighting-aware portrait background replacement. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6452–6462, 2024. 3

    work page 2024

  43. [43]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022. 7, 12

    work page 2022

  44. [44]

    Progressive distillation for fast sampling of diffusion models.arXiv preprint arXiv:2202.00512, 2022

    Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models.arXiv preprint arXiv:2202.00512, 2022. 3

    Pith/arXiv arXiv 2022

  45. [45]

    Follow-your-preference: Towards preference- aligned image inpainting.arXiv preprint arXiv:2509.23082,

    Yutao Shen, Junkun Yuan, Toru Aonishi, Hideki Nakayama, and Yue Ma. Follow-your-preference: Towards preference- aligned image inpainting.arXiv preprint arXiv:2509.23082,

    arXiv

  46. [46]

    Pro- cesspainter: Learning to draw from sequence data

    Yiren Song, Shijie Huang, Chen Yao, Hai Ci, Xiaojun Ye, Jiaming Liu, Yuxuan Zhang, and Mike Zheng Shou. Pro- cesspainter: Learning to draw from sequence data. InSIG- GRAPH Asia 2024 Conference Papers, pages 1–10, 2024. 4

    work page 2024

  47. [47]

    Streamingeffect: Real-time human-centric video effect gen- eration.arXiv preprint arXiv:2605.17019, 2026

    Yiren Song, Cheng Liu, Yuxin Jiang, and Mike Zheng Shou. Streamingeffect: Real-time human-centric video effect gen- eration.arXiv preprint arXiv:2605.17019, 2026

    Pith/arXiv arXiv 2026

  48. [48]

    Vista: Triplet-supervised video style transfer with diffusion transformers.arXiv preprint arXiv:2605.17312, 2026

    Yiren Song, Wangzi Yao, Haofan Wang, and Mike Zheng Shou. Vista: Triplet-supervised video style transfer with diffusion transformers.arXiv preprint arXiv:2605.17312, 2026

    Pith/arXiv arXiv 2026

  49. [49]

    Emo: Emote portrait alive generating expressive portrait videos with audio2video diffusion model under weak conditions

    Linrui Tian, Qi Wang, Bang Zhang, and Liefeng Bo. Emo: Emote portrait alive generating expressive portrait videos with audio2video diffusion model under weak conditions. In European Conference on Computer Vision, pages 244–260. Springer, 2024. 4

    work page 2024

  50. [50]

    Wan: Open and advanced large-scale video generative models.arXiv preprint arXiv:2503.20314, 2025

    Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianx- iao Yang, et al. Wan: Open and advanced large-scale video generative models.arXiv preprint arXiv:2503.20314, 2025. 3

    Pith/arXiv arXiv 2025

  51. [51]

    Unipaint: Unified space-time video inpainting via mixture-of-experts

    Zhen Wan, Chenyang Qi, Zhiheng Liu, Tao Gui, and Yue Ma. Unipaint: Unified space-time video inpainting via mixture-of-experts. InProceedings of the IEEE/CVF Inter- national Conference on Computer Vision, pages 1861–1871,

    work page

  52. [52]

    Tam- ing rectified flow for inversion and editing.arXiv preprint arXiv:2411.04746, 2024

    Jiangshan Wang, Junfu Pu, Zhongang Qi, Jiayi Guo, Yue Ma, Nisha Huang, Yuxin Chen, Xiu Li, and Ying Shan. Tam- ing rectified flow for inversion and editing.arXiv preprint arXiv:2411.04746, 2024. 3

    Pith/arXiv arXiv 2024

  53. [53]

    Videolcm: Video latent consistency model.arXiv preprint arXiv:2312.09109,

    Xiang Wang, Shiwei Zhang, Han Zhang, Yu Liu, Yingya Zhang, Changxin Gao, and Nong Sang. Videolcm: Video latent consistency model.arXiv preprint arXiv:2312.09109,

    Pith/arXiv arXiv

  54. [54]

    LiveEdit: Towards Real-Time Diffusion-Based Streaming Video Editing

    Xinyu Wang, Chongbo Zhao, Fangneng Zhan, and Yue Ma. Liveedit: Towards real-time diffusion-based streaming video editing.arXiv preprint arXiv:2606.26740, 2026. 3

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  55. [55]

    Make-your-video: Customized video generation using textual and structural guidance.IEEE transactions on visualization and computer graphics, 31(2): 1526–1541, 2024

    Jinbo Xing, Menghan Xia, Yuxin Liu, Yuechen Zhang, Yong Zhang, Yingqing He, Hanyuan Liu, Haoxin Chen, Xiaodong Cun, Xintao Wang, et al. Make-your-video: Customized video generation using textual and structural guidance.IEEE transactions on visualization and computer graphics, 31(2): 1526–1541, 2024. 4

    work page 2024

  56. [56]

    Motioncanvas: Cinematic shot design with controllable image-to-video generation

    Jinbo Xing, Long Mai, Cusuh Ham, Jiahui Huang, Anirud- dha Mahapatra, Chi-Wing Fu, Tien-Tsin Wong, and Feng Liu. Motioncanvas: Cinematic shot design with controllable image-to-video generation. InProceedings of the Special Interest Group on Computer Graphics and Interactive Tech- niques Conference Conference Papers, pages 1–11, 2025. 4

    work page 2025

  57. [57]

    Georelight: Learning joint geometrical relighting and reconstruction with flex- ible multi-modal diffusion transformers.arXiv preprint arXiv:2604.20715, 2026

    Yuxuan Xue, Ruofan Liang, Egor Zakharov, Timur Bagaut- dinov, Chen Cao, Giljoo Nam, Shunsuke Saito, Gerard Pons-Moll, and Javier Romero. Georelight: Learning joint geometrical relighting and reconstruction with flex- ible multi-modal diffusion transformers.arXiv preprint arXiv:2604.20715, 2026. 3

    Pith/arXiv arXiv 2026

  58. [58]

    Unified video editing with temporal reasoner

    Xiangpeng Yang, Ji Xie, Yiyuan Yang, Yan Huang, Min Xu, and Qiang Wu. Unified video editing with temporal reasoner. arXiv preprint arXiv:2512.07469, 2025. 3

    Pith/arXiv arXiv 2025

  59. [59]

    Cogvideox: Text-to-video diffusion models with an expert transformer.arXiv preprint arXiv:2408.06072, 2024

    Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiao- han Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer.arXiv preprint arXiv:2408.06072, 2024. 3

    Pith/arXiv arXiv 2024

  60. [60]

    Dilightnet: Fine-grained light- ing control for diffusion-based image generation

    Chong Zeng, Yue Dong, Pieter Peers, Youkang Kong, Hongzhi Wu, and Xin Tong. Dilightnet: Fine-grained light- ing control for diffusion-based image generation. InACM SIGGRAPH 2024 Conference Papers, pages 1–12, 2024. 3

    work page 2024

  61. [61]

    Lumen: Consistent video relighting and harmo- nious background replacement with video generative mod- els.arXiv preprint arXiv:2508.12945, 2025

    Jianshu Zeng, Yuxuan Liu, Yutong Feng, Chenxuan Miao, Zixiang Gao, Jiwang Qu, Jianzhang Zhang, Bin Wang, and Kun Yuan. Lumen: Consistent video relighting and harmo- nious background replacement with video generative mod- els.arXiv preprint arXiv:2508.12945, 2025. 3

    Pith/arXiv arXiv 2025

  62. [62]

    Physg: Inverse rendering with spherical gaus- sians for physics-based material editing and relighting

    Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snavely. Physg: Inverse rendering with spherical gaus- sians for physics-based material editing and relighting. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 5453–5462, 2021. 3

    work page 2021

  63. [63]

    Neural video portrait relighting in real-time via con- sistency modeling

    Longwen Zhang, Qixuan Zhang, Minye Wu, Jingyi Yu, and Lan Xu. Neural video portrait relighting in real-time via con- sistency modeling. InProceedings of the IEEE/CVF interna- tional conference on computer vision, pages 802–812, 2021. 4

    work page 2021

  64. [64]

    Scal- ing in-the-wild training for diffusion-based illumination har- monization and editing by imposing consistent light trans- port

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Scal- ing in-the-wild training for diffusion-based illumination har- monization and editing by imposing consistent light trans- port. InThe Thirteenth International Conference on Learn- ing Representations, 2025. 2, 3, 8, 9, 10, 11

    work page 2025

  65. [65]

    Modeling indirect illumination for inverse rendering

    Yuanqing Zhang, Jiaming Sun, Xingyi He, Huan Fu, Rongfei Jia, and Xiaowei Zhou. Modeling indirect illumination for inverse rendering. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 18643–18652, 2022. 3

    work page 2022

  66. [66]

    Controlvideo: Training-free controllable text-to-video generation

    Yabo Zhang, Yuxiang Wei, XIAOPENG ZHANG, Wang- meng Zuo, Qi Tian, et al. Controlvideo: Training-free controllable text-to-video generation. InInternational Con- ference on Learning Representations, pages 54441–54461,

    work page

  67. [67]

    Lumis- culpt: A consistency lighting control network for video gen- eration

    Yuxin Zhang, Dandan Zheng, Biao Gong, Jingdong Chen, Ming Yang, Weiming Dong, and Changsheng Xu. Lumis- culpt: A consistency lighting control network for video gen- eration. 2024. 3

    work page 2024

  68. [68]

    Motiondirector: Motion customization of text-to-video diffusion models

    Rui Zhao, Yuchao Gu, Jay Zhangjie Wu, David Jun- hao Zhang, Jia-Wei Liu, Weijia Wu, Jussi Keppo, and Mike Zheng Shou. Motiondirector: Motion customization of text-to-video diffusion models. InEuropean Conference on Computer Vision, pages 273–290. Springer, 2024. 4

    work page 2024

  69. [69]

    Trajec- tory consistency distillation.arXiv e-prints, pages arXiv– 2402, 2024

    Jianbin Zheng, Minghui Hu, Zhongyi Fan, Chaoyue Wang, Changxing Ding, Dacheng Tao, and Tat-Jen Cham. Trajec- tory consistency distillation.arXiv e-prints, pages arXiv– 2402, 2024. 3

    work page 2024

  70. [70]

    Light-a-video: Training-free video relighting via progressive light fusion.arXiv preprint arXiv:2502.08590, 2025

    Yujie Zhou, Jiazi Bu, Pengyang Ling, Pan Zhang, Tong Wu, Qidong Huang, Jinsong Li, Xiaoyi Dong, Yuhang Zang, Yuhang Cao, et al. Light-a-video: Training-free video relighting via progressive light fusion.arXiv preprint arXiv:2502.08590, 2025. 2, 3, 8, 9, 10, 11

    Pith/arXiv arXiv 2025

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.