H3 R2VA Style-Transfer Workflow Shared With Full ComfyUI JSON
SIR_NVAX_A_LOT · reddit · 2026-09-07
The author shares an R2VA style-transfer workflow built on MiniMax H3, with a drop-in ComfyUI JSON in the comments.
The setup combines their existing .char character-consistency workflow with style/wardrobe transfer from a previous video: extracting only the ascended goddess form's luminous style from the reference video while explicitly excluding its face, identity, proportions and scene; the subject's photoreal face stays constant throughout. Audio score is generated separately by MiniMax H3 and remuxed in. The prompt demonstrates strong constraint-writing: fixed shot count, permanent damage rules, light-source rules and a per-second timeline.
More from Multimodal
- A shark-as-villain clip made with Seedance 2.5 shows off an original prompt — techhalla · 2026-09-07
- Snorricam H3 LoRA released on HuggingFace and Civitai for extreme POV shots — NebSH83 · 2026-09-07
- GPT-6 Astra makes a kids' cartoon with Higgsfield, edits in Premiere, uploads to YouTube — LexiLove · 2026-09-07
- Minimax-h3 Fine-tuned Video Model Trends on Hugging Face — WarmBloodAban · 2026-09-07
- Magnific releases 2-minute AI animated short 'Slice' with all prompts published — cuenca · 2026-09-07
- Free open-source Chinese AI tool turns one idea into a finished video, 110K GitHub stars — Shruti_0810 · 2026-09-07