# Localized generative motion: measured trial Updated October 19, 2026. We want a street scene that moves naturally between the camera's actual stills. The street should stay where it was. A person shouldn't dissolve into a stripe or return to their original clothes only at the next source frame. The latest experiment narrows generation to a moving foreground crop. It separates the path, body support, source appearance and observed background. That gives the model a smaller job, but it doesn't automatically give it the right gait. ## What changed The previous LTX inpainting trial kept a guided body solid and moving forward. It changed brown and dark red clothing to olive and yellow. The mask also clipped generated limbs. Wan 2.2 VACE accepts a source video, a per-frame edit mask, first and last images and appearance references. We used all of those controls in one bounded trial. White mask pixels permitted generation; black pixels retained conditioning. The mask had extra space around the body and excluded the observed front occluder. The source endpoints had black masks. We generated 81 frames at 30 fps in the crop. The supplied guide was 576 × 384; the returned video was 608 × 416. We recorded that discrepancy rather than quietly stretching the result into the source geometry. We disabled automatic temporal downsampling and additional FILM/RIFE interpolation. The reviewed frames came from the generative stage itself. ## What the frames show I reviewed every raw frame. The person stays brown and opaque, which improves on the LTX clothing drift. The gait still looks too rigid. The final raw frame then changes pose and appearance abruptly. Reinserting an exact source image at that point would preserve the endpoint while hiding none of the discontinuity. This trial is rejected for the main viewer. Its [recipe, raw result, contact sheets and review](wan-vace-localized-benchmark.json) remain available for comparison. ## The next comparison A separate Kling O3 trial conditions on a first and last crop spanning 3.737 seconds of source receipt time. It uses a four-second generation. That small retiming must remain explicit in any result report. All 97 raw frames have now been reviewed. The moving figure generally stays opaque and progresses rightward, but the model makes the nearly stationary figure walk toward the crop edge. The moving figure's appearance also changes, including bright head and shoulder regions and later blending with a red barrier. No full localized composite is approved. The [raw review and exact recipe](kling-localized-benchmark.json) retain this result. Runway Aleph 2 is another applicable route. Its API supports up to five timestamped guidance images, which could place the irregular source observations directly along the clip. Its current request schema has no per-pixel mask, so a separate foreground compositor would still control where generated pixels enter the full scene. A Runway developer account is separate from the existing fal balance. No Runway request has been submitted. ## Continuing scope New paid processing now focuses on Canal × Allen. The Times Square experiments above were already running before that change. We tried to cancel the remaining non-Allen enhancement requests; all ten cancellation responses said those jobs had already completed. We can't claim cancellation savings. The larger goal is still appearance-preserving pose synthesis over an observed background. Higher frame count doesn't solve that on its own. ## Primary references - [Wan 2.2 VACE inpainting API](https://fal.ai/models/fal-ai/wan-22-vace-fun-a14b/inpainting/api). - [Official VACE user guide](https://github.com/ali-vilab/VACE/blob/main/UserGuide.md). - [Kling O3 reference-to-video API](https://fal.ai/models/fal-ai/kling-video/o3/pro/reference-to-video/api). - [Runway's current video editing request types](https://github.com/runwayml/sdk-python/blob/main/src/runwayml/types/video_to_video_create_params.py). - [Runway pricing](https://docs.dev.runwayml.com/guides/pricing/).