☕ One coffee = one happy creator: https://buymeacoffee.com/the_frizzy1
Eight of the nine local MiniMax H3 modes in one workflow, picked from one dropdown. Text to video, image to video, last frame, first+last, references, multiframe, ControlNet pose and Extend a clip. Every mode comes out with sound.
What you need
✔ Custom nodes — None. Core ComfyUI only.
✔ ComfyUI version — 0.39.2 for v2.3.0 (native MiniMax H3, SDPose, Switch / Custom Combo / Math Expression, Trim Audio Duration). The benchmark table below was run on 0.37.4.
✔ VRAM — 12 GB — tested, not estimated. Peak 11.3–11.6 GB
✔ System RAM — 32 GB. It peaks at 17–22 GB, so close your other big apps
✔ Tested on — RTX 3060 12 GB · 32 GB DDR4 · ComfyUI 0.37.4 (Sage attention); v2.3.0 on ComfyUI 0.39.2
✔ Licence — Workflow MIT · MiniMax H3: MiniMax H3 Community License
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.
Required — all 10 files, about 69 GB. Text and image to video only (including last frame, first + last and extend) needs 5 of them, about 42 GB: the FL2VA model, the FL2V turbo LoRA, the text encoder and both VAEs.
diffusion_models/
File: minimax_h3_fl2va_pruned_int8_convrot.safetensors — 21.0 GB (12 GB (tested))
Download:
https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/diffusion_models/minimax_h3_fl2va_pruned_int8_convrot.safetensors
Put it here:
ComfyUI/models/diffusion_models/
File: minimax_h3_ref2va_pruned_int8_convrot.safetensors — 21.0 GB (12 GB (tested))
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_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors — 2.0 GB
Download:
https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/loras/minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors
Put it here:
ComfyUI/models/loras/
File: minimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16.safetensors — 2.0 GB
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/
model_patches/
File: minimax_h3_fun_controlnet_union_pruned_int8_convrot.safetensors — 2.3 GB (ControlNet mode)
Download:
https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/model_patches/minimax_h3_fun_controlnet_union_pruned_int8_convrot.safetensors
Put it here:
ComfyUI/models/model_patches/
Comfy-Org/SDPose — the ControlNet mode's pose models.
checkpoints/
File: sdpose_wholebody_fp16.safetensors — 1.9 GB (ControlNet mode)
Download:
https://huggingface.co/Comfy-Org/SDPose/resolve/main/checkpoints/sdpose_wholebody_fp16.safetensors
Put it here:
ComfyUI/models/checkpoints/
diffusion_models/
File: rt_detr_v4-x-hgnet_fp16.safetensors — 0.1 GB (ControlNet mode)
Download:
https://huggingface.co/Comfy-Org/SDPose/resolve/main/diffusion_models/rt_detr_v4-x-hgnet_fp16.safetensors
Put it here:
ComfyUI/models/diffusion_models/
Modes
One Mode dropdown on the node: T2V (prompt) · I2V (First frame) · Last frame · First + last frame · R2V (Reference 1 + 2, write <Picture 1>, <Picture 2> in the prompt) · Multiframe (references + 3 guide frames at the times you set) · ControlNet pose (Reference 1 + video, the pose is read from it) · Extend a clip.
The inputs start bypassed. Un-bypass only the ones your mode uses (select + Ctrl+B) and keep the rest bypassed. Length in seconds is turned into a valid H3 frame count for you: 5 s = 124 frames. Turbo LoRAs are on by default, 8 steps FL2VA and 4 steps Ref2VA, 20 steps with turbo off. Two LoRA slots sit on the node: strength 0 means the file is never even opened.
Extend with the clip's sound, new in v2.3.0. The clip's last 0.9 s of sound go in with its last 22 frames. In my test the first 0.9 s of the new clip followed the music's loudness almost exactly with the switch on, envelope match 0.961, and not at all with it off, -0.016. After the join H3 makes new sound for your prompt either way: it did not keep the song going. The result begins with those 22 carried-over frames, so trim about 0.9 s when you join clips.
Speed on my RTX 3060
Each mode run once, 5 s clips, turbo on. Wall is queue to finished including model loading, sampling is time in the sampler nodes.
T2V — 0.4 MP, 3 min 02 s wall, 2 min 28 s sampling, 11.3 GB peak vram, 16.9 GB peak ram
I2V (first frame) — 0.4 MP, 3 min 20 s wall, 2 min 46 s sampling, 11.3 GB peak vram, 17.1 GB peak ram
Last frame — 0.4 MP, 3 min 19 s wall, 2 min 46 s sampling, 11.4 GB peak vram, 17.1 GB peak ram
First + last frame — 0.4 MP, 3 min 29 s wall, 2 min 54 s sampling, 11.3 GB peak vram, 16.9 GB peak ram
R2V (references) — 0.4 MP, 2 min 10 s wall, 1 min 33 s sampling, 11.4 GB peak vram, 16.9 GB peak ram
Multiframe — 0.4 MP, 2 min 18 s wall, 1 min 42 s sampling, 11.3 GB peak vram, 16.9 GB peak ram
ControlNet (pose) — 0.4 MP, 4 min 59 s wall, 2 min 43 s sampling, 11.3 GB peak vram, 20.1 GB peak ram
Extend, with the clip's sound — 0.4 MP, 3 min 51 s – 3 min 53 s wall, — sampling, 11.4 GB peak vram, 19.1 GB peak ram
Extend, silent — 0.4 MP, 3 min 31 s – 3 min 33 s wall, — sampling, 11.4 GB peak vram, 19.1 GB peak ram
At 0.9 MP (1280x736) every mode also ran: 6 min 31 s for Multiframe up to 13 min 07 s for ControlNet, peak VRAM 11.6 GB, peak RAM 18.5–21.8 GB. Keep Megapixels at 0.4 or 0.9 on 12 GB. 1.5 MP spills out of VRAM: T2V timed out after 32 min, so stay at 0.9 or below.
What did not work
An input you are not using, left un-bypassed, gives you "Invalid image file" before the run even starts. ComfyUI checks the file of every active loader, also in modes that never read it.
If the mode needs an input and it is still bypassed, the run stops at a "Needs ..." node instead of quietly making a T2V.
Downloaded only the 42 GB set? The Ref2VA, Ref2V LoRA, ControlNet, SDPose and RT-DETR loaders go red. Click them and pick any file you have in that folder, they are never loaded in the FL2VA modes.
Loaders red although you downloaded everything: your models sit in a sub-folder like
diffusion_models/MiniMax-H3/. Re-pick each red loader once.There is no loop node in core ComfyUI yet, so a long video here means running Extend again and again, one clip after another.
The ninth H3 mode, a full character swap, is not in here. It runs on Viggle's H3 finetune and needs a custom node pack, so it stayed its own workflow.
Get the workflow
GitHub: https://github.com/Thefrizzy1/The_frizzy1-Workflows/tree/main/workflows/minimax/h3-ultimate-12gb
Hugging Face mirror: https://huggingface.co/The-frizzy1/The_frizzy1-Workflows/tree/main/workflows/minimax/h3-ultimate-12gb
More videos on this stuff: https://www.youtube.com/playlist?list=PLMztlpsP7Bgg
Description
v2.3.0 adds "Extend: keep the clip's sound", so the last 0.9 s of the clip's audio goes in with its last 22 frames and the join does not jump. Tested on ComfyUI 0.39.2.