Verified four-window run · exact workflow included

A 19-second motion transfer you can inspect and reproduce

A static ivory android adopts the dance of a visually unrelated mechanical performer. The three panels keep the input, driver, and all 305 generated frames visible so motion fidelity, identity stability, and continuity can be judged rather than merely claimed.

Reference

Motion driver

Generated output

Output: 305 frames · 19.0625 sec
Quality gate: Four-window gate passed
Successful path: About 25 min · $0.20
Tested scope: Four generation windows

The successful mask preflight and render cost about $0.20. Total experiment spend was about $0.41 because an earlier completed render was lost when the downloader ignored ComfyUI's temporary output, forcing a rerun.

Download the exact workflow Inspect the advanced deployment

The Gist contains the exact public V4.2 workflow. The hosted route is an advanced technical environment, not a promise that a new input will reproduce this result without SAM2 tuning. Review the current GPU rate before launching; running deployments are billed until stopped.

What this run verifies

A reproducible 305-frame motion-transfer proof

The output follows the unrelated driver’s full-body dance for all 305 frames instead of inventing arbitrary motion.
The three joins between four generation windows were inspected frame-by-frame with no visible seam.
The ivory android identity, bald mechanical head, fitted costume, hands, and feet remain recognizable without flesh, hair, a duplicate performer, cropped feet, or black padding.
A sampler-free SAM2 preflight held across the full driver, so the paid render ran once with zero point corrections.
The exact sanitized workflow, custom-node revisions, model revisions, source trim, and hashes are public.

What it does not verify

Beyond four windows and self-serve remain open gates

This run stops at four windows; where identity drift begins beyond 19 seconds is still unknown.
SAM2 first-frame points require visual validation and may need manual tuning for each new subject or driver.
Driver and target silhouette compatibility materially affects costume and anatomy stability.
The opening push-in comes from the source camera, and this remains an advanced workflow rather than a guaranteed one-click result.

The useful workflow lesson

Validate the mask before paying for diffusion

Prompt changes alone did not fix earlier identity leakage. The reliable improvement came from matching the target and driver silhouettes, then previewing the SAM2 removal mask across the full driver before starting the paid generation step.

1
Prepare the target
Use a full-body character with clear limbs and framing compatible with the driver.
2
Align the window math
Use 77 frames for the first window plus 76 new frames per extension; 305 frames produces four windows without final padding.
3
Inspect SAM2
Preview the entire removal mask and tune first-frame points before generation.
4
Render and validate
Check identity, anatomy, framing, motion fidelity, continuity, media metadata, and cost.

Current product status

The four-window proof and exact downloadable workflow are public. The hosted environment and deployment route exist, but each new input can still require mask tuning and visual QA. Extending beyond four windows is the next duration gate; qualified workflow downloads, deployment interest, and user attempts are the distribution gate.

Download workflow Check your own workflow View licensed driver source