arxiv: 2604.13425 · v1 · submitted 2026-04-15 · 💻 cs.CV

Recognition: unknown

VibeFlow: Versatile Video Chroma-Lux Editing through Self-Supervised Learning

Bin Fu, Jiaying Liu, Pei Cheng, Shuai Yang, Yifan Li

Authors on Pith no claims yet

Pith reviewed 2026-05-10 13:11 UTC · model grok-4.3

classification 💻 cs.CV

keywords video editingself-supervised learningchroma-lux editingvideo relightingcolor editingzero-shotpre-trained modelstemporal coherence

0 comments

The pith

Pre-trained video generators can be adapted via self-supervised perturbations to edit color and lighting in videos without any paired training data.

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 prove that video chroma-lux editing, which changes illumination and color while holding structure and timing steady, does not require expensive supervised training on synthetic pairs. Instead, it shows that the built-in knowledge inside existing video generation models can be accessed and recombined using a disentangled perturbation pipeline that pulls structure from the source video and color-illumination cues from a reference image. A reader would care because this removes the main barrier of creating large labeled datasets and training from scratch, while still supporting many practical tasks such as relighting, recoloring, low-light fixes, day-to-night shifts, and object color changes. The work adds Residual Velocity Fields and a Structural Distortion Consistency Regularization to correct motion discretization errors and keep edits coherent over time.

Core claim

VibeFlow shows that a disentangled data perturbation pipeline on pre-trained video generation models can force adaptive recombination of structure from source videos with color-illumination cues from reference images, producing robust self-supervised disentanglement for chroma-lux editing. Residual Velocity Fields correct discretization errors in flow-based models while Structural Distortion Consistency Regularization enforces structural preservation and temporal coherence, allowing the system to generalize zero-shot across multiple editing tasks without task-specific training or paired data.

What carries the argument

The disentangled data perturbation pipeline, which separates and recombines structure from the source video with color-illumination cues from reference images.

If this is right

The same trained system works zero-shot on video relighting, recoloring, low-light enhancement, day-night translation, and object-specific color editing.
Computational cost drops sharply because no new supervised training or synthetic paired datasets are needed.
Structural and temporal fidelity are maintained through the added Residual Velocity Fields and consistency regularization.
Visual quality remains high while overhead is reduced compared with fully supervised alternatives.

Where Pith is reading between the lines

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

The approach implies that other pre-trained generative models may hold similar latent physical knowledge that could be unlocked with comparable self-supervised pipelines.
Because training resources are no longer required, the method could lower the barrier for smaller teams or individual creators to experiment with video editing tools.
The zero-shot property suggests the framework might extend to new editing tasks beyond those demonstrated, provided suitable reference images are supplied.

Load-bearing premise

Pre-trained video generation models already contain an intrinsic, separable understanding of physical illumination and color that a self-supervised perturbation pipeline can reliably extract and recombine.

What would settle it

Videos edited with the method would show visible structural warping, flickering, or loss of fine detail in scenes with fast motion or complex lighting, or the outputs would fail to match the reference color and light when tested on diverse unseen references.

Figures

Figures reproduced from arXiv: 2604.13425 by Bin Fu, Jiaying Liu, Pei Cheng, Shuai Yang, Yifan Li.

**Figure 1.** Figure 1: Our proposed VibeFlow achieves versatile video color and illumination editing while strictly preserve structural fidelity, enjoying rich applications, including video relighting, recoloring, day-night translation, low-light enhancement, and object-specific color editing. damental challenge of video chroma-lux editing lies in simultaneously satisfying three critical objectives: temporal coherence, photorea… view at source ↗

**Figure 2.** Figure 2: Overall framework of our VibeFlow. We propose a self-supervised training scheme with Structure/Chroma-lux data perturbation pipeline to finetune a video-to-video generation model using LoRA. Without any paired or rendered data, our approach achieves superior video chroma-lux editing with high-fidelity structural preservation. where λ(t) denotes the reweighting scheme based on the timestep t, c indicates au… view at source ↗

**Figure 3.** Figure 3: Residual Velocity-guided Trajectory Rectification. We define the residual velocity field to serve as structural compensation, correcting the editing sampling process iteratively. Trajectory Rectification. During inference, the residual velocity field acts as a geometric anchor to rectify the editing trajectory conditioned on target color/light references. Let V t edit = vθˆ(xt, t, {I i src} N i=1, Iref) d… view at source ↗

**Figure 5.** Figure 5: Visualized Comparison. Our VibeFlow outperforms both specialized methods and general methods in terms of visual quality, temporal consistency and structural fidelity. Datasets. We adopt 60 videos from DAVIS-val [Perazzi et al., 2016] as the evaluation set and 67520 samples from OpenVid-1M [Nan et al., 2024] as the training set. Metrics. We mainly adopt VBench [Huang et al., 2024] to calculate quantitative … view at source ↗

**Figure 6.** Figure 6: Rich Applications. Our VibeFlow can generalize to video low-light enhancement in a zero-shot manner with any specific training/finetuning, while demonstrating superior visual quality. (a) Input Video & Reference Image (Top-left yellow box bounded one) (b) Object Color Editing by our VibeFlow [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗

**Figure 7.** Figure 7: Rich Applications. Our VibeFlow can zero-shot generalize to object color editing, handling complex colors replacement. Notably, the wrinkles pointed by arrows are faithfully preserved, and colors strictly adhere to the reference image. to nighttime requires not only editing on colors (sky from blue to black), but also darken the overall illumination. With the powerful capability learned from disentangle t… view at source ↗

**Figure 9.** Figure 9: Ablation Study. (Compare a&c) The lack of selfsupervised disentangled training and residual velocity leading to severe chaotic eye movements and mismatched background sunlight. (Compare b&c) Employing residual velocity-guided trajectory rectification and consistency training further alleviates the inherent discretization errors and improves fine-grained facial fidelity. fine-grained facial details (e.g… view at source ↗

**Figure 8.** Figure 8: Rich Applications. Beyond basic color-light editing, our VibeFlow also achieves day-night translation, manipulating the colors of sky and illumination in a versatile manner. 4.4 Ablation Study We perform ablation studies on our two key designs in [PITH_FULL_IMAGE:figures/full_fig_p007_8.png] view at source ↗

**Figure 1.** Figure 1: Generalization clarification on hard cases [PITH_FULL_IMAGE:figures/full_fig_p011_1.png] view at source ↗

**Figure 2.** Figure 2: Code example that includes all parameter configurations used in color/light perturbation. Readers can quickly implement a basic VibeFlow-style SSL training paradigm on their own projects by referencing the codes above. Original Video Frame Augmented Ver.1 Augmented Ver.2 Augmented Ver.3 Augmented Ver.4 Augmented Ver.5 [PITH_FULL_IMAGE:figures/full_fig_p012_2.png] view at source ↗

**Figure 3.** Figure 3: Visualization of color/light perturbed results [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗

read the original abstract

Video chroma-lux editing, which aims to modify illumination and color while preserving structural and temporal fidelity, remains a significant challenge. Existing methods typically rely on expensive supervised training with synthetic paired data. This paper proposes VibeFlow, a novel self-supervised framework that unleashes the intrinsic physical understanding of pre-trained video generation models. Instead of learning color and light transitions from scratch, we introduce a disentangled data perturbation pipeline that enforces the model to adaptively recombine structure from source videos and color-illumination cues from reference images, enabling robust disentanglement in a self-supervised manner. Furthermore, to rectify discretization errors inherent in flow-based models, we introduce Residual Velocity Fields alongside a Structural Distortion Consistency Regularization, ensuring rigorous structural preservation and temporal coherence. Our framework eliminates the need for costly training resources and generalizes in a zero-shot manner to diverse applications, including video relighting, recoloring, low-light enhancement, day-night translation, and object-specific color editing. Extensive experiments demonstrate that VibeFlow achieves impressive visual quality with significantly reduced computational overhead. Our project is publicly available at https://lyf1212.github.io/VibeFlow-webpage.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

VibeFlow gives a concrete self-supervised perturbation pipeline plus residual fields for video color-light edits, but the zero-shot claims rest on an unverified disentanglement assumption with no numbers to check.

read the letter

The key point is that this paper shows how to adapt frozen pre-trained video models for chroma-lux editing by feeding them recombined structure from the source video and color-illumination from a reference, then training only on reconstruction. They add residual velocity fields to handle flow discretization and a structural consistency regularizer to keep geometry and timing stable. That specific recipe is the new piece and it avoids the usual supervised paired-data requirement, which is a practical step forward for tasks like relighting or day-night shifts. The public webpage helps too. The main weakness is that the whole thing depends on the pre-trained model already holding separable physical factors that the perturbation can pull out cleanly. Nothing in the description forces that separation beyond the final reconstruction objective, so it could be picking up dataset correlations instead. The abstract talks about impressive results and zero-shot use across five applications but gives no metrics, ablations, or error bars, which leaves the soundness hard to assess. If the full experiments show clear gains over baselines and confirm the factors are actually disentangled, the method would be worth using. Right now the evidence is too thin to know. This is for people building on generative video models who need editing without new labeled data. A serious referee should see the full results and code to decide if the assumption holds or if the gains are real.

Referee Report

3 major / 2 minor

Summary. The paper proposes VibeFlow, a self-supervised framework for video chroma-lux editing that leverages pre-trained video generation models via a disentangled data perturbation pipeline. This pipeline recombines structure from source videos with color-illumination cues from reference images, while introducing Residual Velocity Fields and Structural Distortion Consistency Regularization to mitigate discretization errors in flow-based models and ensure structural/temporal fidelity. The work claims zero-shot generalization to tasks including relighting, recoloring, low-light enhancement, day-night translation, and object-specific editing, without requiring expensive supervised training on paired synthetic data.

Significance. If the central claims hold, the work would be significant for video editing and generative modeling by showing how frozen pre-trained models can be adapted self-supervisedly for physical attribute manipulation, substantially lowering the barrier of paired data collection and training compute. The specific regularization terms target a known limitation in flow-based video models.

major comments (3)

[Abstract] Abstract: the claim that the framework 'unleashes the intrinsic physical understanding' of pre-trained models and achieves zero-shot generalization to diverse tasks is asserted without any quantitative metrics, ablation studies, error bars, or baseline comparisons, which is load-bearing for the central claim of effectiveness and reduced training resources.
[Method] Method description (perturbation pipeline): no explicit loss term, architectural constraint, or auxiliary objective is described that enforces disentanglement between structure and color-illumination factors; training reduces to reconstruction of the recombined target, so success may rely on dataset shortcuts rather than verified access to independent physical factors in the frozen latents.
[Experiments] Experiments section: the contributions of Residual Velocity Fields and Structural Distortion Consistency Regularization are presented as rectifying discretization errors, but without ablations isolating their impact on structural preservation or temporal coherence, it is impossible to assess whether they are load-bearing or merely incremental.

minor comments (2)

[Abstract] The term 'chroma-lux editing' is used without a formal definition or citation to prior literature establishing the scope of the problem.
[Abstract] The project webpage link is provided but no details on released code, models, or evaluation protocols are given in the manuscript, limiting reproducibility assessment.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We address each major comment below, providing clarifications where the manuscript already contains supporting material and outlining revisions where additional evidence or discussion will strengthen the presentation.

read point-by-point responses

Referee: [Abstract] Abstract: the claim that the framework 'unleashes the intrinsic physical understanding' of pre-trained models and achieves zero-shot generalization to diverse tasks is asserted without any quantitative metrics, ablation studies, error bars, or baseline comparisons, which is load-bearing for the central claim of effectiveness and reduced training resources.

Authors: The abstract summarizes the core claims, while the full manuscript (Section 4) presents extensive qualitative results across five tasks, visual comparisons against existing methods, and demonstrations of zero-shot performance without paired training data. We agree that stronger referencing to the experimental evidence would improve the abstract. In the revision we will add a concise sentence directing readers to the quantitative user studies, perceptual metrics, and baseline comparisons reported in the experiments, along with a note on the absence of paired-data training. revision: yes
Referee: [Method] Method description (perturbation pipeline): no explicit loss term, architectural constraint, or auxiliary objective is described that enforces disentanglement between structure and color-illumination factors; training reduces to reconstruction of the recombined target, so success may rely on dataset shortcuts rather than verified access to independent physical factors in the frozen latents.

Authors: The disentanglement arises from the explicit construction of the training targets: each target is formed by taking structure exclusively from the source video and color-illumination exclusively from the reference image. Because the frozen video model must reconstruct this recombined input, it is incentivized to route the two factors through separate pathways in its latent space. We will expand the method section with a dedicated paragraph and accompanying diagram that formalizes this inductive bias and includes feature visualizations showing separation of structure and appearance cues. While an auxiliary loss could be added, the current design already avoids trivial shortcuts by construction. revision: partial
Referee: [Experiments] Experiments section: the contributions of Residual Velocity Fields and Structural Distortion Consistency Regularization are presented as rectifying discretization errors, but without ablations isolating their impact on structural preservation or temporal coherence, it is impossible to assess whether they are load-bearing or merely incremental.

Authors: We acknowledge that the current ablation table does not fully isolate the two proposed components. In the revised manuscript we will add a dedicated ablation study that reports variants with and without Residual Velocity Fields and with and without the Structural Distortion Consistency Regularization. The new table will include quantitative measures of structural fidelity (edge preservation, landmark consistency) and temporal coherence (flow warping error, frame-to-frame consistency scores) together with standard deviations across multiple sequences. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation relies on explicit pipeline design and experimental validation

full rationale

The paper introduces a disentangled perturbation pipeline that constructs training pairs by separating structure (from source video) and color-illumination (from reference), then trains the model via reconstruction on the recombined target. This separation is an explicit input construction rather than a derived or fitted result, and the zero-shot applications follow directly from using the adapted model with new references. No equations, self-citations, or fitted parameters are presented that reduce the central claims (disentanglement or generalization) to definitional equivalence with the inputs. The assumption of intrinsic understanding in the pre-trained model is a modeling premise, not a load-bearing reduction shown by the paper's own logic. The framework is therefore self-contained against its stated objectives.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 2 invented entities

The central claim rests on the unverified domain assumption that pre-trained video models already encode separable physical cues for structure versus illumination/color, plus two newly introduced components whose independent utility is not demonstrated.

axioms (1)

domain assumption Pre-trained video generation models encode disentangleable physical cues for structure, color, and illumination that can be accessed without supervision.
Invoked when the paper states that the framework 'unleashes the intrinsic physical understanding' of these models.

invented entities (2)

Residual Velocity Fields no independent evidence
purpose: Rectify discretization errors inherent in flow-based models.
Introduced to ensure structural preservation and temporal coherence.
Structural Distortion Consistency Regularization no independent evidence
purpose: Enforce rigorous structural preservation and temporal coherence.
Added alongside residual velocity fields to correct flow-model limitations.

pith-pipeline@v0.9.0 · 5511 in / 1513 out tokens · 60434 ms · 2026-05-10T13:11:47.676312+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

44 extracted references · 22 canonical work pages · 11 internal anchors

[1]

arXiv preprint arXiv:2511.06271 , year=

[Bianet al., 2025] Weikang Bian, Xiaoyu Shi, Zhaoyang Huang, Jianhong Bai, Qinghe Wang, Xintao Wang, Pengfei Wan, Kun Gai, and Hongsheng Li. RelightMas- ter: Precise video relighting with multi-plane light images. arXiv:2511.06271,

work page arXiv 2025
[2]

Stable Video Diffusion: Scaling Latent Video Diffusion Models to Large Datasets

[Blattmannet al., 2023a] Andreas Blattmann, Tim Dock- horn, Sumith Kulal, Daniel Mendelevitch, Maciej Kil- ian, 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:2311.15127,

work page internal anchor Pith review arXiv
[3]

Emerging properties in self-supervised vi- sion transformers

[Caronet al., 2021] Mathilde Caron, Hugo Touvron, Ishan Misra, Herv´e J´egou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vi- sion transformers. InProc. Int’l Conf. Computer Vision,

2021
[4]

L-C4: Language- based video colorization for creative and consistent color

[Changet al., 2024] Zheng Chang, Shuchen Weng, Huan Ouyang, Yu Li, Si Li, and Boxin Shi. L-C4: Language- based video colorization for creative and consistent color. arXiv:2410.04972,

work page arXiv 2024
[5]

PerTouch: Vlm-driven agent for personalized and semantic image retouching

[Changet al., 2026] Zewei Chang, Zheng-Peng Duan, Jianxing Zhang, Chun-Le Guo, Siyu Liu, Hyungju Chun, Hyunhee Park, Zikun Liu, and Chongyi Li. PerTouch: Vlm-driven agent for personalized and semantic image retouching. InProc. AAAI Conference of Artifical Intelligence,

2026
[6]

VINO: A uni- fied visual generator with interleaved omnimodal context

[Chenet al., 2026] Junyi Chen, Tong He, Zhoujie Fu, Pengfei Wan, Kun Gai, and Weicai Ye. VINO: A uni- fied visual generator with interleaved omnimodal context. arXiv:2601.02358,

work page arXiv 2026
[7]

Scaling rectified flow transformers for high- resolution image synthesis

[Esseret al., 2024] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas M ¨uller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boe- sel, et al. Scaling rectified flow transformers for high- resolution image synthesis. InProc. IEEE Int’l Conf. Ma- chine Learning,

2024
[8]

arXiv preprint arXiv:2501.16330 (2025)

[Fanget al., 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:2501.16330,

work page arXiv 2025
[9]

FlowPortal: Residual-corrected flow for training-free video relighting and background replace- ment.arXiv: 2511.18346,

[Gaoet al., 2025a] Wenshuo Gao, Junyi Fan, Jiangyue Zeng, and Shuai Yang. FlowPortal: Residual-corrected flow for training-free video relighting and background replace- ment.arXiv: 2511.18346,

work page arXiv
[10]

TokenFlow: Consistent diffu- sion features for consistent video editing.Proc

[Geyeret al., 2024] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. TokenFlow: Consistent diffu- sion features for consistent video editing.Proc. Int’l Conf. Learning Representations,

2024
[11]

Nano banana,

[Google, 2025] Google. Nano banana,

2025
[12]

AnimateDiff: Animate Your Personalized Text-to-Image Diffusion Models without Specific Tuning

https://aistudio. google.com/models/gemini-2-5-flash-image. [Guoet al., 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:2307.04725,

work page internal anchor Pith review arXiv 2023
[13]

Denoising diffusion probabilistic models

[Hoet al., 2020] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. InAdvances in Neural Information Processing Systems,

2020
[14]

LoRA: Low-rank adaptation of large language models

[Huet al., 2022] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. InProc. Int’l Conf. Learning Represen- tations,

2022
[15]

VBench: Comprehensive benchmark suite for video gen- erative models

[Huanget al., 2024] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. VBench: Comprehensive benchmark suite for video gen- erative models. InProc. IEEE Int’l Conf. Computer Vision and Pattern Recognition,

2024
[16]

QuadPrior++: Multi-dimension augmented physical prior for zero-reference illumination enhancement.IEEE Trans- actions on Pattern Analysis and Machine Intelligence,

[Huanget al., 2025] Haofeng Huang, Yifan Li, Wenjing Wang, Wenhan Yang, Ling Yu Duan, and Jiaying Liu. QuadPrior++: Multi-dimension augmented physical prior for zero-reference illumination enhancement.IEEE Trans- actions on Pattern Analysis and Machine Intelligence,

2025
[17]

DeepRemaster: Temporal source-reference attention net- works for comprehensive video enhancement.ACM Trans- actions on Graphics, 38(6):1–13,

[Iizukaet al., 2019] Satoshi Iizuka and Edgar Simo-Serra. DeepRemaster: Temporal source-reference attention net- works for comprehensive video enhancement.ACM Trans- actions on Graphics, 38(6):1–13,

2019
[18]

Vace: All-in-one video creation and editing.arXiv preprint arXiv:2503.07598, 2025

[Jianget al., 2025] Zeyinzi Jiang, Zhen Han, Chaojie Mao, Jingfeng Zhang, Yulin Pan, and Yu Liu. V ACE: All-in-one video creation and editing.arXiv:2503.07598,

work page arXiv 2025
[19]

Taming flow-based i2v models for creative video editing.arXiv:2509.21917,

[Konget al., 2025] Xianghao Kong, Hansheng Chen, Yuwei Guo, Lvmin Zhang, Gordon Wetzstein, Maneesh Agrawala, and Anyi Rao. Taming flow-based i2v models for creative video editing.arXiv:2509.21917,

work page arXiv 2025
[20]

Anyv2v: A tuning-free framework for any video-to-video editing tasks.arXiv preprint arXiv:2403.14468, 2024

[Kuet al., 2024] Max Ku, Cong Wei, Weiming Ren, Harry Yang, and Wenhu Chen. AnyV2V: A tuning- free framework for any video-to-video editing tasks. arXiv:2403.14468,

work page arXiv 2024
[21]

FLUX.1 Kontext: Flow Matching for In-Context Image Generation and Editing in Latent Space

[Labs, 2025] Black Forest Labs. FLUX.1 Kontext: Flow matching for in-context image generation and editing in latent space.arXiv: 2506.15742,

work page internal anchor Pith review arXiv 2025
[22]

COCO-LC: Colorfulness controllable language- based colorization

[Liet al., 2024] Yifan Li, Yuhang Bai, Shuai Yang, and Jiay- ing Liu. COCO-LC: Colorfulness controllable language- based colorization. InProc. ACM Int’l Conf. Multimedia,

2024
[23]

Control color: Multimodal diffusion-based interactive image coloriza- tion.arXiv:2402.10855,

[Lianget al., 2024] Zhexin Liang, Zhaochen Li, Shangchen Zhou, Chongyi Li, and Chen Change Loy. Control color: Multimodal diffusion-based interactive image coloriza- tion.arXiv:2402.10855,

work page arXiv 2024
[24]

Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow

[Liuet al., 2022] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and trans- fer data with rectified flow.arXiv:2209.03003,

work page internal anchor Pith review Pith/arXiv arXiv 2022
[25]

UniLumos: Fast and unified im- age and video relighting with physics-plausible feedback

[Liuet al., 2025a] Ropeway Liu, Hangjie Yuan, Bo Dong, Jiazheng Xing, Jinwang Wang, Rui Zhao, Yan Xing, Wei- hua Chen, and Fan Wang. UniLumos: Fast and unified im- age and video relighting with physics-plausible feedback. arXiv:2511.01678,

work page arXiv
[26]

OpenVid-1M: A Large-Scale High-Quality Dataset for Text-to-video Generation

[Nanet al., 2024] Kepan Nan, Rui Xie, Penghao Zhou, Tiehan Fan, Zhenheng Yang, Zhijie Chen, Xiang Li, Jian Yang, and Ying Tai. Openvid-1m: A large-scale high- quality dataset for text-to-video generation.arXiv preprint arXiv:2407.02371,

work page internal anchor Pith review arXiv 2024
[27]

Perazzi, J

[Perazziet al., 2016] F. Perazzi, J. Pont-Tuset, B. McWilliams, L. Van Gool, M. Gross, and A. Sorkine- Hornung. A benchmark dataset and evaluation method- ology for video object segmentation. InProc. IEEE Int’l Conf. Computer Vision and Pattern Recognition,

2016
[28]

SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis

[Podellet al., 2023] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas M ¨uller, Joe Penna, and Robin Rombach. SDXL: Improving la- tent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952,

work page internal anchor Pith review arXiv 2023
[29]

Qwen-Image Technical Report

[Qwen, 2025] Team Qwen. Qwen-image technical report. arXiv: 2508.02324,

work page internal anchor Pith review arXiv 2025
[30]

Learning transferable visual models from nat- ural language supervision

[Radfordet al., 2021] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agar- wal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from nat- ural language supervision. InProc. IEEE Int’l Conf. Ma- chine Learning,

2021
[31]

High-resolution image synthesis with latent diffusion models

[Rombachet al., 2022] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. InProc. IEEE Int’l Conf. Computer Vision and Pattern Recognition,

2022
[32]

RAFT: Re- current all-pairs field transforms for optical flow

[Teedet al., 2020] Zachary Teed and Jia Deng. RAFT: Re- current all-pairs field transforms for optical flow. In Proc. European Conf. Computer Vision,

2020
[33]

Videox-fun,

[Videox-fun, 2025] Videox-fun. Videox-fun,

2025
[34]

Wan: Open and Advanced Large-Scale Video Generative Models

https: //github.com/aigc-apps/VideoX-Fun. [Wanet al., 2025] Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianxiao Yang, et al. Wan: Open and advanced large-scale video generative models.arXiv preprint arXiv:2503.20314,

work page internal anchor Pith review Pith/arXiv arXiv 2025
[35]

Image quality assessment: From error visibility to structural similarity.IEEE Trans- actions on Image Processing,

[Wanget al., 2004] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: From error visibility to structural similarity.IEEE Trans- actions on Image Processing,

2004
[36]

Video models are zero-shot learners and reasoners

[Wiedemeret al., 2025] Thadd ¨aus Wiedemer, Yuxuan Li, Paul Vicol, Shixiang Shane Gu, Nick Matarese, Kevin Swersky, Been Kim, Priyank Jaini, and Robert Geirhos. Video models are zero-shot learners and reasoners. arXiv:2509.20328,

work page internal anchor Pith review arXiv 2025
[37]

Rerender a video: Zero-shot text- guided video-to-video translation

[Yanget al., 2023] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. Rerender a video: Zero-shot text- guided video-to-video translation

2023
[38]

CogVideoX: Text-to-Video Diffusion Models with An Expert Transformer

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

work page internal anchor Pith review arXiv
[39]

ImgEdit: A Unified Image Editing Dataset and Benchmark

[Yeet al., 2025] Yang Ye, Xianyi He, Zongjian Li, Bin Lin, Shenghai Yuan, Zhiyuan Yan, Bohan Hou, and Li Yuan. ImgEdit: A unified image editing dataset and benchmark. arXiv: 2505.20275,

work page internal anchor Pith review arXiv 2025
[40]

Lumen: Consistent video relighting and harmonious background replacement with video generative models.arXiv:2508.12945,

[Zenget al., 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 harmonious background replacement with video generative models.arXiv:2508.12945,

work page arXiv 2025
[41]

Adding conditional control to text-to-image dif- fusion models

[Zhanget al., 2023] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image dif- fusion models. InProc. IEEE Int’l Conf. Computer Vision and Pattern Recognition,

2023
[42]

Scaling in-the-wild training for diffusion-based illumination harmonization and editing by imposing con- sistent light transport

[Zhanget al., 2025] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Scaling in-the-wild training for diffusion-based illumination harmonization and editing by imposing con- sistent light transport. InProc. Int’l Conf. Learning Rep- resentations,

2025
[43]

VCGAN: Video colorization with hy- brid generative adversarial network.IEEE Transactions on Multimedia, 25:3017–3032,

[Zhaoet al., 2022] Yuzhi Zhao, Lai-Man Po, Wing-Yin Yu, Yasar Abbas Ur Rehman, Mengyang Liu, Yujia Zhang, and Weifeng Ou. VCGAN: Video colorization with hy- brid generative adversarial network.IEEE Transactions on Multimedia, 25:3017–3032,

2022
[44]

Light-A-Video: Training-free video relighting via progressive light fusion

[Zhouet al., 2025] Yujie Zhou, Jiazi Bu, Pengyang Ling, Pan Zhang, Tong Wu, Qidong Huang, Jinsong Li, Xiaoyi Dong, Yuhang Zang, Yuhang Cao, Anyi Rao, Jiaqi Wang, and Li Niu. Light-A-Video: Training-free video relighting via progressive light fusion. InProc. Int’l Conf. Computer Vision,

2025