Edited endpoint images make the target of an image-to-video transition explicit and can improve transition completion, but also preserve context in the source image's layout. Full-frame endpoint conditioning turns this preserved context into requirements on the video's final state, potentially inducing temporal artifacts. We measure discontinuity with Hard-Cut (HC), late background reversal with Background Motion Reversal (BMR), and motion suppression with Dynamic Degree (DD). These artifacts appear across five video models spanning training-free and learned endpoint conditioning, in model-specific forms. We study Partial, a selective conditioning scheme that constrains only an edit-relevant region estimated from image-editor attention and leaves the remaining terminal content to the video model. Across three backbones on TC-Bench-I2V, Partial reduces temporal artifacts relative to full-frame conditioning, with no significant difference in Transition Completion Ratio (TCR); on Wan2.2-Turbo, uniformly weakening conditioning does not reproduce these reductions. A donor-context intervention keeps full-frame conditioning and replaces only the context outside the edit-relevant region with the final frame of a Partial video. It lowers HC and BMR on HunyuanVideo-1.5 and lowers BMR while restoring DD on Wan2.2-Turbo, with no significant difference in TCR, so constraint extent alone does not account for these artifacts. Together, these findings suggest that edited endpoints should serve as partial future goals, although edit relevance does not guarantee future validity. Code and data will be released at https://github.com/wannagetacat/Edited-Endpoints-I2V.
Figures & tables
Figure 1: Edited endpoints preserve context that need not specify the video’s future. Full-frame endpoint conditioning constrains both requested appearance and preserved context. Partial constrains only estimated edit-relevant support M0 from image-editor attention (blue outline); washed-out positions retain the video prediction. Source frame: DAVIS bus [ 32 ] , via FiVE-Bench [ 20 ] .
Figure 2: Partial reduces temporal artifacts relative to Full. Paired differences with 95% CIs over 103 prompts and five seeds; open markers are not significant. Tables 1 and C1 give levels and intervals.
Figure 3: Terminal content, not only constraint extent, affects temporal artifacts. (a) The donor-context endpoint keeps the edited endpoint inside M0 (blue outline) and uses the final frame of a paired Partial rollout outside it (orange tint). Original and Donor-context Full share the Full projection and the recipient seed; a cross-seed donor comes from another seed of the same prompt. (b) Donor-context minus Original Full with paired 95% CIs over 103 prompts and five seeds; open markers are not significant; gray ticks mark Partial minus Original Full. (c) HunyuanVideo-1.5, prompt 28, seed 22004: native frames and consecutive-frame CLIP similarity; the Original Full discontinuity between frames 114 and 117 is absent under Donor-context Full.
Figure 4: Estimated edit-relevant support does not guarantee a future-valid terminal constraint. Wan2.2-Turbo, purple-bus example of Fig. 1 (case 90001, outside TC-Bench), seed 22001. (a) Saved M0 outline (blue) and token-pooled weights (tint). (b) Partial: the boxed sign is absent at native frames 108 and 112, then reappears at 114. Identical regions are enlarged below.
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
Figure A1: Estimated edit-relevant support. Source images, edited endpoints, and GMM–Ising masks for TC-Bench prompts 0, 5, and 82. Blue marks support; percentages give constrained area before token pooling.
Figure A2: Endpoints from three editors. Source images and edited endpoints for TC-Bench prompts 0, 103, and 105. FLUX.2 and Qwen-Image-Edit use the final-frame prompt in Appendix A ; ChronoEdit uses the original transition prompt to match its formulation of editing as video generation.
Analysis
Population and reference
Main backbone comparisons
103 prompts × 5 seeds per condition; Table 1 .
Wan2.2-Turbo spatial controls
103 prompts × 5 seeds per support, paired with the main Wan2.2-Turbo Partial (TCR 42.91%).
HunyuanVideo-1.5 donor-context intervention
Original Full reuses the main cohort. Donor P regenerates main Partial with identical settings; TCR 44.66% versus 44.47% reflects run-to-run variation.
HunyuanVideo-1.5 cross-seed donors
O, same-seed D, and P are reused; 515 new Dcross videos use a donor-seed assignment fixed before generation.
Wan2.2-Turbo donor-context intervention
O, D, N, and P are new 515-video runs sharing one noise stream per prompt–seed. P reproduces main Partial’s HC, DD, and BMR exactly; TCR 42.72% versus 42.91% reflects re-judging variation. O regenerates Full with this noise stream and stays within sampling variation of main Full on all four metrics (TCR 43.30% versus 41.94%, HC 6.49 versus 6.47, DD 65.44% versus 66.80%, BMR 0.600 versus 0.599; no paired difference is significant).
Table B1: Appendix populations and references. Counts refer to paired prompt–seed positions unless stated otherwise.
Figure C1: Completion requires covering the edit-relevant support. Wan2.2-Turbo, paired Δ TCR relative to Partial with 95% CIs over 103 prompts and five seeds; open markers are not significant. (a) Supports with exactly ∣M0∣ cells covering 0–75% of M0 , the rest placed outside it; 100% is Partial. (b) Supports placed without reference to M0 , which cover 8–10% of it by chance. The dashed line marks Endpoint-free.
Condition
Inside M0
Outside M0
Support
O: Original Full
Original latent
Original latent
Full
D: Donor-context Full
Original latent
Re-encoded donor final RGB
Full
Dcross : Cross-seed donor
Original latent
Re-encoded final RGB of another seed’s donor
Full
N: Native-latent Full
Original latent
Native donor tail latent
Full
P: Partial donor
Original latent
Video prediction retained
M0
Appendix
Table D1: Donor-context construction. O/D/ Dcross /N all retain Full support. Content inside M0 always comes from the original edited endpoint.
Figure F1: A residual failure of fixed support. (a) The edited endpoint preserves an occluding bus-stop sign; blue shows M0 , and the red box defines the crop. (b) Identical image-coordinate crops at all eight original frame indices. Red borders mark the sign’s reappearance in Partial and random release; attention-ranked and adaptive release avoid it in this example. The release variants are exploratory. All rows use Wan2.2-Turbo, purple-bus case 90001 and seed 22001, as in Fig. 4 ; frame intervals are unequal.
Unified conditional image generation requires a single model to preserve different aspects of a visual input while producing a target image that follows textual guidance. Pretrained video models provide a promising foundation by encoding visual transformations across time, yet existing approaches typically assign the condition and target to fixed temporal positions. We observe that condition-target relationships vary substantially across generation tasks, while video models naturally encode visual transformations at different temporal scales. This motivates us to model conditional generation through input-adaptive virtual temporal relations. We propose FrameForge, which captures two complementary aspects of such relations. Mixture of Temporal Experts (MoTE) models how far the target lies from the condition by dynamically selecting temporal RoPE experts with different virtual intervals. Chain of Temporal Reasoning (CoTR) models how the condition is transformed into the target through a latent intermediate state, with a reasoning-bottleneck attention mask encouraging information propagation through this pathway. Experiments on DreamBench, COCO2017, and GEdit-Bench show that FrameForge achieves superior performance in subject-driven generation, canny-to-image generation, and image editing among the compared methods.
Building a capable video editor remains significantly harder than a video generator: editing requires (source, instruction, edited) triplets that are prohibitively expensive to annotate and difficult to synthesize at scale, whereas image editing has already reached maturity with millions of such pairs readily available. In this work, we introduce VINCIE-NExT, a unified framework that transfers editing capability from images to videos through in-context visual demonstrations, alleviating the need for large-scale paired video editing data. VINCIE-NExT decomposes video editing into a structured chain of composable sub-tasks (Video -> Image -> Image -> Video), routing editing intent through the image domain and enabling scalable joint training from heterogeneous image and video corpora under a unified diffusion objective. An image editing pair, synthesized by the model or supplied by the user, is prepended as an in-context visual demonstration that serves as a spatial appearance blueprint for every output frame. To ground appearance edits across the interleaved context, we introduce a novel position encoding that links image demonstrations and video frames in a shared spatial coordinate system, enabling pixel-faithful propagation of appearance changes to every output frame. Chain-of-Editing further provides principled test-time scaling: by executing the sub-task chain as progressive diffusion stages, editing quality can be improved by investing additional compute without retraining. Comprehensive experiments on OpenVE-Bench demonstrate the state-of-the-art performance across diverse editing categories, with ablations confirming the effectiveness of each component.
Leigang Qu, Feng Cheng, Ziyan Yang +7
National University of Singapore · ByteDance Seed · University of Science and Technology of China
Diffusion Transformers (DiTs) have demonstrated remarkable scalability and quality in image and video generation, prompting growing interest in extending them to controllable generation and editing tasks. However, compared to the image counterparts, progress in video control and editing remains limited, mainly due to the scarcity of paired video data and the high computational cost of training video diffusion models. To address this issue, in this paper, we propose a video-free tuning framework termed ViFeEdit for video diffusion transformers. Without requiring any forms of video training data, ViFeEdit achieves versatile video generation and editing, adapted solely with 2D images. At the core of our approach is an architectural reparameterization that decouples spatial independence from the full 3D attention in modern video diffusion transformers, which enables visually faithful editing while maintaining temporal consistency with only minimal additional parameters. Moreover, this design operates in a dual-path pipeline with separate timestep embeddings for noise scheduling, exhibiting strong adaptability to diverse conditioning signals. Extensive experiments demonstrate that our method delivers promising results of controllable video generation and editing with only minimal training on 2D image data. Codes are available https://github.com/Lexie-YU/ViFeEdit.
Ruonan Yu, Zhenxiong Tan, Zigeng Chen +2
National University of Singapore · Shanghai Jiao Tong University