MiniMax H3 Reference-to-Video Character Swap Tested: Motion Kept, Likeness Drifts
NextDiffusion · reddit · 2026-10-21
NextDiffusion shares day-one tests of character swap with MiniMax H3: replacing a video character via a character-sheet reference image while preserving original motion, scene, and sound, with a full ComfyUI tutorial and workflow downloads.
- Known limitations: character likeness drifts, details may blend with the original person, scene cuts cause timing drift; extra prompting helps consistency
- Results are presented as uncherry-picked; tutorial and workflows are freely available
Related event: MiniMax H3 Character Swap Tests Show Promise with Some Drift(3 posts)→
More from Multimodal
- ByteDance's Dreamina Image 4.0 tests Seedream 4.0 with 6-image fusion and true 4K output — Aiden_Tech_Ai · 2026-10-21
- Flux Klein's reference latent node excels at style transfer — can Qwen 2.1 do it? — Inevitable_Pen9043 · 2026-10-21
- AI-generated character IP demo: one mascot across posters, merch and conference scenes — eyal_zuri · 2026-10-21
- Cinematic AI fantasy romance short 'Alpha Meik & Luna Martina' released — Soundangel-Martina · 2026-10-21
- AcademiaSD adds Fibo 2.0 LoRA training with NF4 quantization, cutting VRAM from 16GB to 5.2GB — AcademiaSD · 2026-10-21
- "We went deeper and found more sky": stunning AI-generated megastructure sci-fi art — piotrbinkowski · 2026-10-21