More references aren’t a face-swap guarantee.

The finding

Should you use fast restyling or a more detailed replacement setup?

Keep the options separate. H1 is the practical speed choice in these examples. HM4 can improve some detail and background behavior, but it is much slower and still leaks source clothing or misses likeness.

Two inputs shown from ten saved video outputs and four previews. H1 versus HM4 is a whole-recipe comparison.

What we would choose

Present H1 as fast restyling and HM4 as an experimental detail option. Do not sell either as a reliable face-only swap or exact celebrity replacement.

Where it still falls short

• HM4 also changes the base model checkpoint, sampling schedule, steps and adapters. Four photos alone cannot be credited with the visible differences.
• The actual archived HM4 test used one appearance anchor plus three crops from different real photos. Current hosted HM4 input choices may differ; it is not an exact one-click replay.
• Four keyframe previews were saved, but keyframe propagation never rendered. It is not a proven third replacement option.

What changed

Time & cost, without the guesswork.

The campaign’s recorded Studio worker debit was $20.4551 for ten videos, four previews and startup/idle/save time. Per-output charges were not invoiced. Do not divide that total evenly or treat engine time as end-to-end waiting.

H1 and HM4 used the same shared H3 worker in this campaign. The exact device class is not reconstructed here; engine timings are not a cross-GPU benchmark.

Inspect the setup

Runtime identity

Archived experimental runtime sha256:5de427d73f1c472f1aaf22e472d543d36e58f33894c8bfea30036df42d6909a9

H1 model checkpoint

minimax_h3_ref2va_pruned_int8_convrot.safetensors

HM4 model checkpoint

minimax_h3_hybrid_fl2va_ref2va_b20-49-int8.safetensors

Reference encoder

qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors

Seed

1414213562

Other saved options

H4, HQ4 and HN4 also completed. More detail or additional views did not establish a universal likeness fix.

Hardware and timing limits

H1 and HM4 used the same shared H3 worker in this campaign. The exact device class is not reconstructed here; engine timings are not a cross-GPU benchmark.

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.

Includes four archived H1/HM4 API graph templates, credits and model requirements. Private runtime nodes and fresh input bindings are required; standalone execution has not been qualified.