Recent advances in image editing have enabled fine-grained and highly controllable visual modifications. However, achieving similar levels of controllability in video editing remains an open challenge, particularly for precise local edits, dynamic temporal effects, and consistent lighting across frames.In this paper, we propose EditDistill, a framework that distills image-level visual transformations from edits to guide video editing.Specifically, given a source video and an edit prompt, we first generate an edited version of the initial video frame using an off-the-shelf image editing model. We then distill the visual transformation of this original-edited image pair into a compact representation termed an edit descriptor.This descriptor is injected into a video diffusion model through our lightweight fusion module, which steers the denoising process toward the desired edit. To align the video trajectory with the demonstrated image edit while preserving motion and layout, we train with edit-descriptor alignment and DINO structural consistency losses, and derive an equivalent inference-time guidance formulation under flow matching.Extensive experiments demonstrate that our method achieves strong performance on standard benchmarks and produces high-quality video editing results across diverse scenarios. For more details and videos, please refer to the anonymous webpage: https://yu-shaonian.github.io/EditDistill-Web/.
Figures & tables
Figure 1: EditDistill distills an image edit pair into a compact transformation descriptor and injects it into a video diffusion model, enabling local object edits, stylization, background replacement, and global relighting while preserving source motion and temporal coherence.
Figure 2: EditDistill encodes the visual transformation between image editing pair into a compact edit descriptor, and combines the edit descriptor alignment constraint with the DINO structural similarity loss to distill the image editing capability into the denoising trajectory of video editing.
Figure 3: Overview of EditDistill . Stage I encodes the transformation between an image pair (Is,Ie) into the edit descriptor d . Stage II keeps the Wan2.1-I2V-14B video diffusion frozen and trains the inject modules to inject d into selected transformer blocks by lightweight feature fusion modules steering the diffusion denoising toward the desired edit to produce a temporally coherent video Ve from the video Vs .
Figure 4: Descriptor attention. Cross-attention maps show that local edits concentrate on the regions that need to change, while global edits such as relighting or stylization produce broader but still structured responses. The descriptor behaves as a transformation operator rather than a fixed mask or copied target image.
Method
CLIP-E ↑
CLIP-T ↑
DINO-sim ↑
Temp. ↑
Warp ↓
FVD ↓
InstructPix2Pix ( Brooks et al., 2023 )
0.221
0.288
0.534
0.892
7.15
1284.7
TokenFlow ( Geyer et al., 2024 )
0.196
0.275
0.668
0.948
4.21
1052.3
AnyV2V ( Ku et al., 2024 )
0.137
0.251
0.591
0.856
6.94
1485.2
RF-Solver-Edit ( Wang et al., 2025b )
0.224
0.286
0.654
0.926
4.58
996.3
MOFA-Video ( Niu et al., 2024 )
0.182
0.271
0.621
0.943
4.83
1024.1
Prompt-only Wan2.1 ( Wang et al., 2025a )
0.158
0.291
0.612
0.951
3.84
982.4
Table 1: Quantitative comparisons. All methods run at matched 480×832 and 81 -frame experimental settings. For evaluation metrics, CLIP-E and CLIP-T measure edit fidelity; DINO-sim measures preservation of unedited regions; Temp. and Warp measure temporal coherence; FVD measures perceptual video quality. Best results are in bold and second best results are underlined .
Method
SubC ↑
BkgC ↑
MoS ↑
AesQ ↑
DyaD ↑
InstructPix2Pix ( Brooks et al., 2023 )
0.842
0.871
0.924
0.429
0.612
TokenFlow ( Geyer et al., 2024 )
0.946
0.951
0.972
0.514
0.524
AnyV2V ( Ku et al., 2024 )
0.812
0.768
0.912
0.431
0.512
RF-Solver-Edit ( Wang et al., 2025b )
0.928
0.939
0.967
0.571
0.628
MOFA-Video ( Niu et al., 2024 )
0.932
0.945
0.974
0.568
0.718
Prompt-only Wan2.1 ( Wang et al., 2025a )
0.921
0.943
0.944
0.562
0.681
Table 2: Comparisons of video editing consistency. SubC: subject consistency; BkgC: background consistency; MoS: motion smoothness; AesQ: aesthetic quality; DyaD: dynamic degree. Best results are in bold ; second best results are underlined .
Figure 5: Qualitative comparison. EditDistill performs both local and global edits while keeping the main subject, background, and motion stable. In contrast, other methods may miss the desired edit, unintentionally change nearby regions, or produce unstable results across frames.
Figure 6: Relighting comparison. The distilled descriptor captures a global lighting transformation and guides coherent relighting across frames while preserving scene geometry and motion.
Variant
Trainable params
CLIP-E ↑
DINO-sim ↑
Warp ↓
FVD ↓
Cross-attention only
312MB
0.243
0.692
3.41
921.4
FiLM only
187MB
0.218
0.704
3.55
938.2
Additive only
154MB
0.196
0.681
3.78
967.5
All branches (ours)
480MB
0.245
0.738
2.97
852.1
Table 3: Ablation study of the edit feature injection module. The three branches are complementary: cross-attention is strongest for local routing, FiLM improves global appearance control, and the additive branch contributes a stable residual edit bias.
Figure 7: Stage I edit descriptor model. The Delta VAE encoder extracts a stochastic descriptor from the source and edited VAE latents by combining shared features, their explicit feature difference, and a multi-scale encoder. The Delta VAE decoder receives only the source latent and descriptor, and reconstructs the edited latent through descriptor-conditioned residual blocks with self- and cross-attention.
Figure 8: Frame-wise descriptor attention. The same edit descriptor re-localizes to the relevant moving regions across different video frames.
Figure 9: Qualitative ablation. Removing individual branches or alignment losses weakens either edit fidelity or temporal preservation, matching the quantitative trends in Section 5.3 .
Video editing aims to modify input videos according to user intent. Recently, end-to-end training methods have garnered widespread attention, constructing paired video editing data through video generation or editing models. However, compared to image editing, the high annotation costs of video data severely constrain the scale, quality, and task diversity of video editing datasets when relying on video generative models or manual annotation. To bridge this gap, we propose LIVE, a joint training framework that leverages large-scale, high-quality image editing data alongside video datasets to bolster editing capabilities. To mitigate the domain discrepancy between static images and dynamic videos, we introduce a frame-wise token noise strategy, which treats the latents of specific frames as reasoning tokens, leveraging large pretrained video generative models to create plausible temporal transformations. Moreover, through cleaning public datasets and constructing an automated data pipeline, we adopt a two-stage training strategy to anneal video editing capabilities. Furthermore, we curate a comprehensive evaluation benchmark encompassing over 60 challenging tasks that are prevalent in image editing but scarce in existing video datasets. Extensive comparative and ablation experiments demonstrate that our method achieves state-of-the-art performance. The source code will be publicly available.
Weicheng Wang, Zhicheng Zhang, Zhongqi Zhang +6
Nankai University · Kuaishou Technology · Pengcheng Laboratory +1
Video editing spans diverse editing paradigms, yet achieving high-quality instruction-guided and subject-guided editing within a single unified framework remains challenging. We introduce EditVid, a training-free framework combining sparse causal memory for local coherence, correspondence-based post-attention token injection for long-range identity preservation, and soft latent blending for edit locality. The same framework supports instruction-guided and reference-guided edits, including style transfer, attribute modification, object insertion, part-level editing, and subject replacement. On FiVE, EditVid achieves 78.16 FiVE-Acc, compared with 58.95 for the strongest evaluated training-free baseline, while obtaining competitive results on IVEBench. A user study further shows a 51.8% overall preference for EditVid over 7 competing methods.
Adheesh Sunil Juvekar, Onkar Kishor Susladkar, Kiet A. Nguyen +6
Recent image editing systems have achieved impressive semantic understanding, visual fidelity, and instruction-following ability, while video editing remains substantially more difficult and costly. In this paper, we present a simple alternative to end-to-end video editing: instead of training a monolithic video editor, we transform a strong image editor into a video editor through anchor-based generation. Our key insight is that video editing can be decomposed into two subproblems: editing a sparse set of keyframes and propagating those edits across time. Based on this observation, we propose Anchor-based Video Editing (AVE), a two-stage framework in which a powerful image editor first performs composed editing on selected keyframes, and a motion-guided image-to-video diffusion model then generates the final video by treating the edited keyframes as fixed anchors. This design directly inherits the strengths of modern image editors while avoiding expensive end-to-end video editing training. Experiments on IVEBench and VIE-Bench show that AVE achieves strong performance in instruction following, temporal consistency, and content fidelity. Further ablations reveal that final video editing quality is strongly correlated with the quality of the image editor, suggesting that future progress in video editing may come from stronger image editing foundations and lightweight transfer to video. Code is available at https://github.com/wangf3014/AVE.