Bringing legacy video back to contemporary screens does not necessarily make it look contemporary. Earlier imaging pipelines often leave color casts, muted chroma, weak con- trast, and tonal imbalance, while correcting these character- istics still relies heavily on expert colorists and scales poorly to long, multi-shot videos. We formulate long video color modernization, a task that updates the chromatic and tonal appearance of legacy color footage while preserving scene content and coherence across frames and shots. We introduce a scalable, training-free framework that jointly modernizes grids of temporally adjacent keyframes using a pretrained text-guided image editing model, producing coherent con- temporary references without paired supervision. Compact shot-specific color transforms then propagate these adjust- ments throughout each shot, preserving spatial structure and temporal stability. We conduct a multidimensional evaluation on diverse legacy videos spanning five decades. We combine a human evaluation of perceived modernness with quantita- tive measures of content fidelity, temporal stability, percep- tual quality, and efficiency, and introduce Source-Weighted Inter-shot Grade Consistency (SW-IGC) to assess grade co- herence across shots. On our evaluation set, our frame- work achieves the highest perceived modernness while re- maining competitive across complementary measures, open- ing a path toward the systematic study of long video color modernization. Code is available at https://github.com/jiwoomood/ Video_Color_Modernization.git
Figures & tables
Figure 1: Overview of our long video color modernization framework. We first partition the input video into shots and select a representative keyframe from each shot. We then jointly modernize temporally adjacent keyframes. Finally, we estimate a shot-specific color mapping from each original–modernized keyframe pair and apply it to all frames in the corresponding shot.
Figure 2: Distribution of contemporary color preference rankings over 975 complete rankings. Each bar shows the proportions of Rank 1, Ranks 2–3, and Ranks 4–8, with the mean rank reported on the right.
Method
Edge-SSIM ↑
CDC ↓
SW-IGC ↓
DeQA ↑
Q-Align Q ↑
Q-Align A ↑
FPS ↑
Original
0.966
0.56
0.19
2.73
3.16
2.32
–
End-to-End Video Editing
CogVideoX-5B
0.439
0.77
–
2.32
2.55
1.92
–
Wan2.1-VACE-1.3B
0.535
1.37
–
1.92
1.96
1.63
–
Modular Color Propagation
Neural Preset
0.886
0.56
7.72
2.73
3.11
2.34
820
Table 1: Complementary quantitative comparison on 100 videos from the Condensed Movies Dataset. The unedited source video (Original) is shown in gray for reference and excluded from ranking. CDC is reported in units of 10−2 . SW-IGC is omitted for end-to-end methods since reliable source–output correspondence is unavailable. FPS denotes the propagation throughput on an NVIDIA H200 GPU (see Appendix B.4 ).
Figure 3: Qualitative comparison of video modernization methods on representative frames.
Strategy
Invocations
SW-IGC ↓
Independent
N
7.23
Independent + post-hoc smoothing
N
4.46 †
Packing (clustered)
⌈N/4⌉
5.87
Packing (consecutive)
⌈N/4⌉
5.90
Table 2: Ablation of keyframe modernization strategies. N denotes the number of shots in each video, and the packing strategies process up to four keyframes per model invocation. † Post-hoc smoothing directly regularizes the shot-specific mappings evaluated by SW-IGC.
Figure 4: Qualitative comparison of inter-shot consistency strategies for keyframe modernization. Independent modernization exhibits inter-shot color inconsistency, while post-hoc smoothing produces unnatural color distortions. The two joint strategies produce more coherent keyframe appearances across shots.
Color mapping operator
SWD ↓
Edge-SSIM ↑
Diagonal
0.821
0.884
Histogram
0.490
0.856
MKL
0.742
0.883
MKL + roll-off
0.748
0.885
Table 3: Ablation of shot-specific color mapping operators. Each operator is fitted and evaluated on the same source keyframe and its modernized counterpart.
Figure 5: Qualitative comparison of shot-specific color propagation operators. (A) Results obtained by applying each fitted mapping to its source keyframe. Top: Diagonal fails to reproduce the target grade, whereas Histogram introduces local color artifacts. Bottom: The per-channel mappings fail to reproduce the cross-channel target grade and produce strong color distortions, whereas MKL closely matches it. (B) Results after propagating the fitted mapping to a brighter frame: hard clipping saturates highlights, while roll-off preserves their tonal variation.
Appendix figures & tables9 assets
Supplementary material from the paper’s appendix.
Appendix
Configuration
Value
Keyframe Extraction
Shot detection
detector
TransNet V2
boundary threshold
0.1
Shot merging
boundary encoder
DINOv2-S
Appendix
Table 4: Pipeline configurations used for all experiments.
Stage
Device
Cost
Shot detection
CPU
∼ 24 s / min of video
Shot merging
GPU
∼ 0.9 s / shot
Keyframe selection
GPU
∼ 13 s / shot
Image-level modernization
GPU
∼ 14 s / grid
Color propagation
CPU
∼ 80 FPS
Appendix
Table 5: Per-stage processing cost, pooled over ten videos from our dataset.
Method
Mean Rank ↓
Rank-1 ↑
Top-3 ↑
Ours
2.313
37.9%
81.7%
CAP-VSTNet
3.015
17.6%
66.3%
Video Color Grading
3.559
12.9%
49.0%
Original
3.619
13.3%
45.1%
Neural Preset
4.028
11.0%
38.6%
SA-LUT
5.461
5.8%
12.8%
Appendix
Table 6: Human evaluation of modern color preference over 975 complete rankings (65 participants × 15 videos). Mean Rank uses all placements from 1 to 8; Rank-1 and Top-3 denote the percentage of rankings in the corresponding range.
Metric
Value
Videos
15
Methods per video
8
Unordered method pairs per video ( (28) )
28
2AFC queries per pair (forward + reversed order)
2
Total VLM queries
840
Complete swapped-order comparisons
420
Appendix
Table 7: Order-swap stability of the VLM-based two-alternative forced-choice (2AFC) color judge. For each unordered method pair, we query the judge twice: once in the original A/B order and once with the two candidates swapped. A comparison is consistent only when both queries select the same underlying method after mapping the A/B labels back to method identities.
Figure 6: Average pairwise absolute differences between the VLM scores of the eight evaluated methods across 15 videos.
Figure 7: Comparison of keyframe selection methods. Each row shows the keyframes selected by different methods for the same shot; failure cases are outlined in red.
Figure 8: Instructions presented to participants in the user study.
Figure 10: Additional qualitative comparisons between our method and the evaluated baselines on diverse frames from the evaluation set (Part 1).
Figure 11: Additional qualitative comparisons between our method and the evaluated baselines on diverse frames from the evaluation set (Part 2).
Long autoregressive video generation faces a fundamental memory challenge: with a finite attention window, a model must decide which information from an ever-expanding history to retain. Existing methods organize memory temporally, preserving recent frames while compressing or discarding older ones. We instead propose RECAP-Forcing, organizing memory by appearance novelty. A long video is not merely a sequence of frames, but an evolving cast of subjects, objects, and scenes whose identities must remain consistent over time. We organize memory by retaining the KV cache associated with newly appearing content--such as entering subjects, disoccluded regions, and newly introduced scenes--at the moment it first becomes visible, prioritizing novelty over recency. Memory should scale with the amount of newly introduced content, rather than with video length. This appearance-indexed memory makes long-range consistency an explicit property of the memory structure. Our framework unifies two mechanisms under this single principle. At the beginning of a video, when all visible content is novel, an attention sink preserves the initial scene. As the video evolves, an optical-flow-based novelty bank extends the same principle by selectively retaining newly revealed content. As a training-free inference method with no additional learnable parameters, RECAP-Forcing consistently improves visual quality and semantic fidelity across multiple strong baselines and outperforms existing memory methods.
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.
Yifan Li, Pei Cheng, Bin Fu +2
Wangxuan Institute of Computer Technology, Peking University · Tencent
Video (camera) trajectory editing aims to synthesize new videos that follow user-defined camera paths while preserving scene content and plausibly inpainting previously unseen regions, upgrading amateur footage into professionally styled videos. Existing VTE methods struggle with precise camera control and long-range consistency because they either inject target poses through a limited-capacity embedding or rely on single-frame warping with only implicit cross-frame aggregation in video diffusion models. To address these issues, we introduce a new VTE framework that 1) explicitly aggregates information across the entire source video via a hybrid warping scheme. Specifically, static regions are progressively fused into a world cache then rendered to target camera poses, while dynamic regions are directly warped; their fusion yields globally consistent coarse frames that guide refinement. 2) processes video segments jointly with their history via a history-guided autoregressive diffusion model, while the world cache is incrementally updated to reinforce already inpainted content, enabling long-term temporal coherence. Finally, we present iPhone-PTZ, a new VTE benchmark with diverse camera motions and large trajectory variations, and achieve state-of-the-art performance with fewer parameters.
Zhihao Shi, Kejia Yin, Weilin Wan +5
McMaster University · University of Toronto · The University of Hong Kong +3