More seconds. The same scene?
An eight-second source became 16- and 24-second videos. The original audio survived; new footage was silent.
One source, two lengths
The original sound was preserved in the test file. This public preview is silent; the generated tail adds no sound.
Inspect the join below: the synchronized comparison covers the shared first 16 seconds.
Inspect the join · first 16s comparison
Compare One source, two lengths
Inspect movement closely
Both videos stay on the same timeline at every speed; this does not retime the character’s moves. Frame buttons seek by 41.7 ms on the output grid; browser decoding and a different source frame rate can affect the displayed frame.
Loading the comparison videos… Open the result file
Silent publication previews. This pair compares the shared interval; full-length files are below. Audio findings refer to preserved original test files.
What to look for At eight seconds the generated continuation begins. Inspect hands, the garment and face through the join and later tail. The original music ends with the original footage.
Original video, full results & downloads
Shorter sections: lower warm cost, imperfect appearance
Compare Shorter sections: lower warm cost, imperfect appearance
Inspect movement closely
Both videos stay on the same timeline at every speed; this does not retime the character’s moves. Frame buttons seek by 41.7 ms on the output grid; browser decoding and a different source frame rate can affect the displayed frame.
Loading the comparison videos… Open the result file
Silent publication previews. This pair compares the shared interval; full-length files are below. Audio findings refer to preserved original test files.
What to look for Compare only new footage after 8s, especially turns and the join at 12s. Stage and broad motion continue; face, hands and fabric still distort. One attempt, not general quality parity.
Original video, full results & downloads
Extra steps did not establish a quality improvement
Compare Extra steps did not establish a quality improvement
Inspect movement closely
Both videos stay on the same timeline at every speed; this does not retime the character’s moves. Frame buttons seek by 41.7 ms on the output grid; browser decoding and a different source frame rate can affect the displayed frame.
Loading the comparison videos… Open the result file
Silent publication previews. This pair compares the shared interval; full-length files are below. Audio findings refer to preserved original test files.
What to look for Same stage source, prompt, seed and canvas. Inspect generated faces, hands and fabric after 8s; extra sampling did not convincingly resolve the defects.
Original video, full results & downloads
AI-generated outputs are identified in each comparison. Public previews omit source audio where present; visual streams are unchanged. Original test files and audio measurements remain preserved. No performer or photographer endorses ModelPilot.
The finding
What do our saved longer-video results actually preserve?
We can show completed extensions, but extra seconds are not the same as guaranteed continuity. These deliveries keep the source prefix and its audio. Hands, faces and clothing can drift in the generated tail. The public previews are silent: original-audio retention describes the preserved test files and measurements, not a soundtrack distributed on this page.
Two completed October 28 deliveries, with earlier failed and partial mechanism comparisons kept in the finding. October 31: one matched-source 20-step and one two 4-second visual Relay candidate, one attempt each; original14-step result reused.
What we would choose
Use Extend for a short continuation you can review. Keep the original audio over the source footage and explain that added footage has no new soundtrack. Do not promise long, seamless scenes from these examples.
Where it still falls short
- Earlier relay-versus-Director work produced one full output and one partial output; two cases did not render. It did not establish an overall winning mechanism.
- The saved clips do not prove flawless anatomy, garment continuity or lip sync. A complete file is not a quality pass.
- These two completed deliveries are illustrations, not a matched relay-versus-Director quality comparison.
What changed
| Choice | Baseline / first option | Candidate / second option |
|---|---|---|
| Source prefix | 8 seconds, original audio | Same saved source |
| Additional footage | +8 seconds | +16 seconds |
| Delivered length | 16 seconds | 24 seconds |
| Added soundtrack | Silent after original ends | Silent after original ends |
Time & cost, without the guesswork.
| Saved run | Recorded processing | Recorded amount | Scope |
|---|---|---|---|
| 8s source → 16s result | 7m 57s | $1.3813 | Warm worker; 31m 32s queue |
| 8s source → 24s result | 30m 55s | $3.4775 | Cold worker; 1m 25s queue |
| 20-step visual Relay · +8s | 14m42s | $2.1345 | Cold; one stage clip |
| Two 4-second visual Relay sections · +8s | 6m45s | $1.1729 | Warm; one stage clip |
| Fixed original appearance · stage +8s | 12m00s | $1.6798 | Cold B200;16s total |
| Moving appearance · landscape +8s | 3m41s | $0.6402 | Warm B200;28.5s total |
| Fixed original appearance · landscape +8s | 6m50s | $0.2679 | Cold RTX PRO6000 96GB;28.5s total |
Historical Studio charges for the complete jobs. More footage, warm reuse and queue time all differ. These numbers are not today’s quote or an equal-work mechanism benchmark. The RTX follow-up debit missed paid startup time; it is not a valid expected price for a future run.
NVIDIA B200 in the saved completed Extend run detail; the +8s job reused the +16s worker.
Inspect the setup
- Frame accounting
- Original 8s = 192 frames at 24fps; outputs 384 and 576 frames
- Source dimensions
- 942 × 1374; model canvas 896 × 1312
- Audio evidence
- Original-audio alignment measured on the prefix; no generated tail soundtrack. Measurement does not establish lip sync.
- Earlier comparison
- Source/Director/relay examples included held source frames for comparison; a held frame is not source motion beyond the original clip.
- Hardware and timing limits
- NVIDIA B200 in the saved completed Extend run detail; the +8s job reused the +16s worker.
- October 31 quality finding
- 20 steps did not show a convincing face/hand/clothing gain. Two 4-second sections retained turns, with remaining distortions; one source cannot establish equivalent quality.
- October 31 warm comparison
- 14-step single 8s:7m57s/$1.3813; two4s:6m45s/$1.1729 (about 15% lower complete warm time and charge). Both B200; startup and queue separate.
- October 31 20-step run
- 14m42s/$2.1345 including cold startup; sampling 569.883s versus saved 14-step 420.529s. More work did not establish better quality.
- Availability distinction
- These visual Relay quality candidates remain private experiments. The current new-draft default is Director-style native-guide22, not the saved visual Relay control.
- Original-appearance follow-up status
- Three completed files are preserved: stage fixed-original16s ($1.6798), landscape moving-reference28.5s ($0.6402), and landscape fixed-original28.5s on RTX PRO6000 96GB ($0.2679). Turns continue, but fixed appearance did not consistently repair hair, faces or hands. The new follow-up files are not distributed as public previews on this page.
- November 1 priced GPU observation
- The final RTX96 job completed in6m50s from allocation, or7m28s after Generate, with a confirmed$0.2679 Studio debit. Its engine took170.395s versus134.212s for the saved moving-reference B200 job. Hardware, conditioning and cold/warm state differ; this is not a controlled hardware-speed or quality-parity benchmark. The provider day report confirms$0.2883 for this exact pod. The historical Studio debit missed112 seconds of allocated startup and under-collected about2¢ versus supplier cost. A future-only clock correction is prepared locally; the old charge will not change. Do not use that debit as an expected future price.
- November 1 quality decision
- Keep the single8-second saved control as the stronger hairstyle example. The4+4 options remain experiments: motion survives, but hair shape pops/inflates and face/hand artifacts remain. A lower hourly machine rate is a separate choice from the generation recipe.
Study settings summary; not an importable ComfyUI workflow. Requires the private ModelPilot runtime/nodes and separately licensed model weights.
Download study filesIncludes recorded settings, media hashes, credits and license notices. A complete executable graph is not included for this study.
Sources, attribution & publication status
Source credits, original licenses and adaptation notices are recorded below. Results are AI-generated research demonstrations, not endorsements. Model use and output redistribution remain subject to the MiniMax H3 license and any required separate authorization.
All videos on this page are saved outputs, not new generations. Settings and conclusions were transcribed from the original experiment receipts. No customer accounts, private batch identifiers or credentials are included in the downloadable summary.
- Patricia Stephan, TAF German Bellydance National Championship 2024 — Dance — Shark1989z. Excerpt from original 25s onward; resampled/prepared for the saved test; audio omitted from public previews. AI-generated extensions are adaptations, not new footage of Patricia Stephan. Copyright permission record: Wikimedia VRT ticket 2024112010013196.. License recorded at acquisition.
Source credits and changes accompany each download. Read the media notices. Public previews omit source audio; the original test files remain preserved. ModelPilot has received authorization for its H3 use and works. This does not grant rights to third-party source material.