Edited Endpoints as Partial Goals for Image-to-Video Generation

Which parts of an edited endpoint should constrain a video's future?

Sihyun Kim1 Sanghyun Jo1,2,† Kyungsu Kim1,†

1Seoul National University2OGQ

†Corresponding authors

Paper (coming soon) Code (coming soon)

Code and data will be released upon publication.

An edited endpoint changes the bus to purple while preserving the source scene. Full conditioning constrains the whole terminal frame; Partial constrains only the edit-relevant support.
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 (Perazzi et al., CVPR 2016) via FiVE-Bench (Li et al., ICCV 2025), licensed under CC BY-NC 4.0; resized and edited.
Key insight

An edited endpoint carries the requested change, but it also keeps the rest of the scene in the source image's layout. Imposing the whole frame turns that preserved context into requirements on the video's final state, a spatial over-specification that shows up as discontinuities, late background reversals, and motion suppression. Imposing only the edit-relevant support keeps completion comparable to full-frame conditioning while reducing these artifacts.

5 models
training-free and learned alike show significant temporal artifacts under full-frame endpoints, in model-specific forms
−38%
HC (−37.5%) and BMR (−38.8%) on HunyuanVideo-1.5 when only the preserved context is replaced, at unchanged full-frame support
35–68%
lower BMR with Partial than with Full on three backbones, with TCR within 2.1 pp of Full
+12.2 pp
Dynamic Degree when Frame Guidance's loss is restricted to M0, back to the Endpoint-free level with the completion gain kept

Abstract

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.

Key findings

01

Terminal content matters beyond constraint extent

With full-frame support unchanged, the donor-context intervention lowers HC and BMR on HunyuanVideo-1.5, and lowers BMR while restoring DD on Wan2.2-Turbo.

02

Support area and placement yield distinct trade-offs

On Wan2.2-Turbo, edit-relevant placement favors completion at matched area; shrinking support lowers HC, whereas enlarging it raises BMR. Tested global softening does not reproduce Partial's artifact reductions.

03

Partial reduces temporal artifacts at comparable completion

Across three backbones, Partial mitigates temporal artifacts relative to full-frame endpoint conditioning at comparable TCR. Restricting Frame Guidance to the same support likewise keeps its completion gain and removes its motion suppression.

Full-frame endpoints pull the background back

Both videos turn the bus purple. Under Full, the video must end exactly on the edited endpoint, including its unchanged background, so the scene returns to the first frame's layout. Under Partial, only the bus is constrained and the camera keeps following it.

Wan2.2-Turbo, same source image, edited endpoint, prompt, and seed for both videos.
Frames of the purple-bus video under Full and Partial, and background motion over time: Full reverses in the second half, Partial does not.
Background Motion Reversal on the purple-bus example. (a) Frames of the two videos above. (b) Background motion outside M0, projected on its early direction and normalized: Full reverses in the second half (BMR 0.95), Partial does not (BMR 0.00). BMR is the share of second-half along-axis motion that runs backward.

Terminal content, not only constraint extent

Full-frame endpoint conditioning and Partial differ in both the imposed terminal content and the constraint extent. The donor-context intervention keeps the edited endpoint inside M0, substitutes the final frame of a paired Partial rollout outside it, and constrains the entire terminal frame with the same full-frame projection.

Donor-context construction, paired differences against Original Full, and an example where the Full discontinuity disappears.
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). (b) Donor-context minus Original Full with paired 95% CIs over 103 prompts and five seeds; open markers are not significant; ticks mark Partial minus Original Full. (c) HunyuanVideo-1.5: the Original Full discontinuity between frames 114 and 117 is absent under Donor-context Full.

Partial versus full-frame conditioning

Paired differences with 95% confidence intervals for TCR, HC, DD, and BMR on three backbones.
Partial reduces temporal artifacts relative to Full. (a) Partial minus Full and (b) Partial minus Endpoint-free, as paired differences with 95% CIs over 103 TC-Bench-I2V prompts and five seeds; open markers are not significant.

Edit-relevant support from image-editor attention

Source images, edited endpoints, and estimated support masks for three prompts.
Estimated edit-relevant support. Source images, edited endpoints, and GMM–Ising masks for three TC-Bench-I2V prompts. Blue marks support; percentages give the constrained area.

Limitation: edit relevance is not future validity

A preserved bus-stop sign disappears and then reappears near the end of a Partial video.
Estimated edit-relevant support does not guarantee a future-valid terminal constraint. An endpoint-preserved sign is absent at frames 108 and 112 under Partial, then reappears at frame 114. Source frame credited in the first figure.

BibTeX

@article{kim2026edited,
  title   = {Edited Endpoints as Partial Goals for Image-to-Video Generation},
  author  = {Kim, Sihyun and Jo, Sanghyun and Kim, Kyungsu},
  journal = {arXiv preprint},
  year    = {2026}
}