☕ One coffee = one happy creator: https://buymeacoffee.com/the_frizzy1
Your video plus one picture of a character gives you the same clip with the person swapped: same moves, same timing, the clip's own sound. Higgsfield's Genjutsu idea run on your own PC, so nothing is uploaded and nothing blocks the clip. The whole clip goes through in 5 s passes with ComfyUI's built-in loop nodes.
What you need
✔ Custom nodes — None. Core ComfyUI only.
✔ ComfyUI version — 0.39 or newer (MiniMax H3 plus the Start Loop / End Loop nodes)
✔ VRAM — 12 GB — tested, not estimated. Peaked at 11892 MiB
✔ System RAM — 32 GB, peaked at 23.1 GB
✔ Tested on — RTX 3060 12 GB · 32 GB DDR4 · ComfyUI 0.39.2
✔ Licence — Workflow MIT · MiniMax H3 and the Character Swap LoRA: MiniMax H3 Community License, read it before any commercial use
Models to download
Click a file name to download it. Put it in the folder shown, inside ComfyUI/models/. Every file name was read from the workflow .json. If you already run my H3 Ultimate workflow you only need the LoRA.
Required — 6 files. The five H3 files are the same ones H3 Ultimate uses.
diffusion_models/
File: minimax_h3_ref2va_pruned_int8_convrot.safetensors — 21.0 GB
Download:
https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/diffusion_models/minimax_h3_ref2va_pruned_int8_convrot.safetensors
Put it here:
ComfyUI/models/diffusion_models/
loras/
File: minimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16.safetensors — 2.0 GB (Turbo (4 steps))
Download:
https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/loras/minimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16.safetensors
Put it here:
ComfyUI/models/loras/
text_encoders/
File: qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors — 15.7 GB
Download:
https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors
Put it here:
ComfyUI/models/text_encoders/
vae/
File: minimax_h3_video_vae_int8_convrot.safetensors — 2.8 GB
Download:
https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/vae/minimax_h3_video_vae_int8_convrot.safetensors
Put it here:
ComfyUI/models/vae/
File: minimax_h3_audio_vae_fp32.safetensors — 0.6 GB
Download:
https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/vae/minimax_h3_audio_vae_fp32.safetensors
Put it here:
ComfyUI/models/vae/
akatz-ai/MiniMax-H3-Character-Swap-LoRA — this is the file that does the swap.
loras/
File: h3_character_swap_pro4500_1000.safetensors — 0.2 GB (the swap)
Download:
https://huggingface.co/akatz-ai/MiniMax-H3-Character-Swap-LoRA/resolve/main/h3_character_swap_pro4500_1000.safetensors
Put it here:
ComfyUI/models/loras/
Settings
All six loaders and every setting are on one main node. If your files sit in a subfolder like MiniMax-H3/, pick them there once.
Who to replace — one sentence, e.g.
big man in the black and gold patterned outfit. It goes into the Character Swap LoRA card's prompt what it doesSeconds — 0 = the whole clip in 5.2 s passes, every pass continuing from the last 22 frames of the one before; a number = only the start what it does
Megapixels — the time knob. 0.4 MP fits 12 GB, more spills into system RAM what it does
Turbo — on = 4 steps with the Ref2V turbo LoRA, off = 20 steps what it does
Keep the original sound — puts the clip's own audio back on what it does
Best input: a character sheet with a front view, a back view and the face.
Speed on my RTX 3060
0.4 MP, Turbo on, whole clip.
wheelbarrow → penguin — 6.7 s length, 2 passes, 11 min 56 s (v1.0.0's own graph: 12 min 14 s) time
van exit → raccoon — 7.0 s length, 2 passes, 10 min 53 s (v1.1.0's own graph: 11 min 46 s) time
Detstvo dance → brown dog — 9.4 s length, 2 passes, 12 min 25 s time
snowball fight → cartoon guy — 8.3 s length, 2 passes, 12 min 18 s time
wrestler flip → penguin — 8.3 s length, 2 passes, 11 min 25 s time
dance → dog — 9.8 s length, 3 passes, 17 min 27 s time
backbend → raccoon — 5.2 s length, 1 passes, 5 min 37 s time
5 s on its own: 9 min 55 s with the Character Swap LoRA, 7 min 31 s without it, 35 min 10 s with 20 steps.
This is slow, and that is the headline, not a footnote: about 11 to 12 minutes for a 7 second clip, and 35 min 10 s if you turn Turbo off and run 20 steps.
What did not work
VRAM peaked at 11892 MiB and system RAM at 23.1 GB, so 12 GB and 32 GB are both near the limit. If RAM fills up, lower Megapixels; 0.4 MP is what fits 12 GB.
From the LoRA's own card: shots of 4 to 5 s work best, which is why the loop runs in 5 s passes. One person at a time.
Hard cuts in the clip confuse it.
If it swaps the wrong person or nobody, your "who to replace" sentence has to describe what that person wears.
Swap in characters, not real people's faces.
Get the workflow
Hugging Face mirror: https://huggingface.co/The-frizzy1/The_frizzy1-Workflows/tree/main/workflows/minimax/h3-character-swap-12gb
More videos on this stuff: https://www.youtube.com/playlist?list=PLMztlpsP7Bgg
Description
v1.1.0 puts every setting and all six model loaders on one main node, so the inputs, the settings and the model files sit in one place. The README has no changelog, so this is read off the v1.0.0 / v1.1.0 files.


