Smaller motion inputs. Shorter waits.
Two complete comparisons cut recorded run charges by 26–38%. Faces and hands still need work.
Studio dance
Compare Studio dance
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 The costume and broad dance stay similar. Inspect the turn around 8.5 seconds: head direction, feet and hands still vary.
Original video, full results & downloads
Indoor dance
Compare Indoor dance
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 The smaller candidate retains the teal costume and short hair near the end. Face direction and expression remain different from the input.
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
Can a smaller motion reference save money without spoiling the result?
Yes, in these two dances. Reducing the motion-video input kept broadly similar movement while lowering time and recorded run charges. The character photo and final output size were not what we made smaller.
Two complete paired comparisons from six attempted cases. Single seed; sampled visual assessment.
What we would choose
Keep the smaller motion-reference setup as a named option for the tested dances. Keep Pose Efficient v1 frozen. Do not advertise identical quality for every clip.
Where it still falls short
- The jumping-jacks control failed to start; the full-reference pole control delivered only eight seconds. Those cases cannot prove a full-clip saving.
- Hair, face direction, feet and hand contact remain imperfect. A smaller input is a speed tradeoff, not an identity fix.
- Physical workers and cache state differed. These are two observations, not a universal speed guarantee.
What changed
| Choice | Baseline / first option | Candidate / second option |
|---|---|---|
| Motion-video reference | Full input resolution | 384 × 736; indoor case 256 × 512 |
| Generation | 8-step turbo + 4-step upscale | Unchanged |
| Body guidance | RTMW body pose | Unchanged |
| Output resolution | Original tested canvas | Unchanged |
Time & cost, without the guesswork.
| Saved run | Recorded processing | Recorded amount | Scope |
|---|---|---|---|
| Studio dance · full input | 26m 39s | $3.9283 | Run charge |
| Studio dance · smaller | 16m 32s | $2.4339 | 38.0% lower charge |
| Indoor dance · full input | 16m 43s | $2.4574 | Run charge |
| Indoor dance · smaller | 12m 17s | $1.8054 | 26.5% lower charge |
Historical recorded run charges, not a current quote. Time is the saved render/download observation, not a prediction of queue wait. Build and shared hosting costs are excluded.
Requested NVIDIA B200 for the recorded full/smaller comparisons; physical workers and cache state differed. The later Compact option targets RTX PRO 6000 (96GB).
Inspect the setup
- Model family
- MiniMax H3 with pose guidance
- Pinned seed
- 1414213562
- Segmentation
- 192-frame windows with 22-frame overlap
- Original configuration
- Pose Efficient v1, frozen October 14
- Related option
- Pose Compact v1: 256 × 512 motion input on the measured 96GB GPU; two later section-time comparisons improved 25–28%. Not the same timing scope as the rows above.
- Hardware and timing limits
- Requested NVIDIA B200 for the recorded full/smaller comparisons; physical workers and cache state differed. The later Compact option targets RTX PRO 6000 (96GB).
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.
- Studio dance source — Andres Daza. Prepared and excerpted for the saved test; no performer endorsement. License recorded at acquisition.
- Indoor dance source — ROMAN ODINTSOV. Prepared excerpt; fictional generated appearance reference. 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.