MiniMax H3 reference-to-video character swap tested, full ComfyUI workflow released
NextDiffusion · reddit · 2026-10-21
NextDiffusion shares day-one tests (no cherry-picking) of character swap in video using MiniMax H3 reference-to-video plus a LoRA, with a full ComfyUI tutorial and workflow downloads. A character reference sheet replaces the person while preserving original motion, scene, and audio. Known limits: likeness drift, details blending with the original person, and timing drift at scene cuts; extra prompting helps consistency.
Related event: MiniMax H3 Video Character Swap Tested: Motion Preserved, Faces Still Drift(2 posts)→
More from Multimodal
- Blender + Wan 3 recreates one New York street corner across 400 years in striking split-screen — angierova · 2026-10-21
- Prof. reveals full script of 'Gaussians Unite' trailer made with Codex + Blender — CSProfKGD · 2026-10-21
- Fully AI-generated café UGC video fools viewers, creator shares every prompt — socialwithaayan · 2026-10-21
- Fibo 2.0 LoRA Trained in 42 Minutes for $2 Puts Plush Dragon Into Famous Paintings — SirToki · 2026-10-21
- HeyGen's Professional Voice Clone uses 2 hours of audio to preserve breaths, pauses and laughs — rohanpaul_ai · 2026-10-21
- Ask an AI to draft architectural drawings of Google's Gradient Canopy — the results look straight out of a blueprint archive — fofrAI · 2026-10-21