A bigger encoder didn’t fix the face.
Two matched comparisons kept the movement. Neither showed a convincing identity improvement.
Florence reference · four-second pole turn
Compare Florence reference · four-second pole turn
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 Gross rotation and framing remain. The face stays approximate, with hand/grip and cloth artifacts. More encoder precision did not earn promotion.
Original video, full results & downloads
Zendaya reference · eight-second dance
Compare Zendaya reference · eight-second 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 dance remains recognizable. Check face likeness, finger shapes and skirt/leg geometry: the intended identity improvement is not convincing.
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
Does changing reference-encoder precision make likeness more accurate?
Not in these two cases. Switching the reference encoder from NVFP4 AWQ to official INT8 did not convincingly improve likeness. The pole turn and studio dance survived; faces, fingers and clothing remained imperfect.
Two encoder-only completed comparisons using saved H1 controls; two planned Extend cases did not render and are excluded here.
What we would choose
Keep the existing H1 default. Do not pay for a precision upgrade on the assumption it fixes identity. A better identity-conditioning mechanism needs separate evidence.
Where it still falls short
- Two examples do not prove INT8 can never help. They do reject a claimed improvement for these saved inputs.
- Cold/warm execution and runtime differences prevent a controlled speed or cost conclusion.
- The deeper cause of source-face dominance is not proven by this negative result.
What changed
| Choice | Baseline / first option | Candidate / second option |
|---|---|---|
| Reference encoder | Qwen3-VL NVFP4 AWQ | Official Qwen3-VL INT8 |
| Input, prompt and seed | Saved H1 control | Held to the saved case |
| Motion and generation recipe | H1 | H1; not a new pose system |
| Observed likeness | Approximate | Still approximate |
Time & cost, without the guesswork.
| Saved run | Recorded processing | Recorded amount | Scope |
|---|---|---|---|
| Florence · saved control | 8m 17s | $0.9164 | Studio job; queue 1m 16s |
| Florence · INT8 | 7m 51s | $0.9760 | Cold job; queue 16s |
| Zendaya · saved control | 3m 07s | Not itemized | Historical engine time only |
| Zendaya · INT8 | 3m 31s | $0.6100 | Warm job; queue 6m 39s |
The two new Studio debits total $1.586. Earlier native checks/app build add an estimated $0.3127 separately; their invoice is unconfirmed. Provider credit movement is not another GPU charge. Timings are explicitly different scopes.
New renders used NVIDIA B200 and shared one physical worker. Old and new runtime/cold-warm states differ; the timing table does not establish an encoder-speed improvement.
Inspect the setup
- Seed
- 1414213562
- Delivered frames
- 96 for Florence; 192 for Zendaya, both at 24fps
- Execution evidence
- Real Studio quote/Start, same-worker handoff, close/return and matching downloaded output hashes
- Current disposition
- Private experiment; not promoted as a quality upgrade
- Hardware and timing limits
- New renders used NVIDIA B200 and shared one physical worker. Old and new runtime/cold-warm states differ; the timing table does not establish an encoder-speed improvement.
Read the archived studio-dance H1 prompt
This is the eight-second Zendaya example. The pole example has its own duration and appearance description; the detailed recipe also assigns additional reference roles.
subject_definitions: The adult performer in <Video 1> supplies movement and scene only. <Picture 1> supplies the replacement identity, outfit and hairstyle. summary: The output is an edited version of <Video 1>, replacing its sole adult performer with the long auburn hair, recognizable facial structure, red sequined bodice with black edging, side cutouts, long black skirt and black pumps of <Picture 1>. retention_analysis: Keep the source camera, background, light, timing, number of people, head and body orientation, limbs, hands, contacts and occlusions. Keep the target costume fully covering the body. detailed_description: Follow only the actual movements visible in <Video 1>, at their original times. Do not add turns, poses or back views that the source does not make. Hair and clothing follow the motion. Do not transplant photograph backgrounds or introduce cuts. One continuous eight-second shot. overall_soundscape: Silent output; no generated audio. non_diegetic_music: None.
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.
- Pole-performance source — Pexels User Flerko. Prepared excerpt; no performer endorsement. License recorded at acquisition.
- Florence Pugh appearance anchor — Frank Sun / WikiPortraits. Original photo; displayed contained. License recorded at acquisition.
- Zendaya appearance anchor — © Glenn Francis, www.PacificProDigital.com. Original photo; displayed contained. 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.