Download
1 variant available
This checkpoint includes a config file, download and place it along side the checkpoint.
940 1 2 3 4 5 6 7 8 90 1 2 3 4 5 6 7 8 9
(8)
Sep 21, 2026
MiniMax H3
210 1 2 3 4 5 6 7 8 90 1 2 3 4 5 6 7 8 9
290 1 2 3 4 5 6 7 8 90 1 2 3 4 5 6 7 8 9
MiniMax H3 is licensed by MiniMax under the MiniMax H3 Community License Agreement. That agreement’s Applicable Territory excludes the European Union, the United Kingdom, the Republic of Korea and the United States of America. Your use of H3 and of any H3 derivative is subject to that agreement and its Acceptable Use Policy.
MiniMax H3
MiniMax H3 Ref2VA — Audio-Driven Turbo (Image + Audio to Video)
A ComfyUI workflow for MiniMax H3 that drives generation with your own reference audio instead of relying on the model's own invented sound. Built on H3's ref2va (reference-to-video/audio) task — the only H3 task type whose node exposes an audio-reference input — so a real audio file (dialogue for lip-sync, an SFX cue like a splash, ambience, etc.) can be handed to the model alongside a reference image and threaded into the generated clip's timing.
Built entirely from stock ComfyUI core nodes — no custom node packs required.
What it does
- Takes a reference image and a reference audio clip, plus a text prompt, and generates a video conditioned on both via MiniMaxH3ReferenceToVideo.
- Auto-appends a fixed audio-retention instruction to your prompt <Audio 1>: fully_copy — <Audio 1> is reused 1:1 as the target video's complete final audio track) via a StringConcatenate node, so the model is always told to preserve your audio as the literal final track rather than treating it as a loose style cue — you don't have to remember to type this yourself.
- Turbo LoRA switcher: pick any turbo LoRA from a dropdown, and a StringContains-based logic chain reads the filename for "8step" or "4step" and automatically sets the matching step count — so the step count can't silently drift out of sync with whichever LoRA file you actually selected.
- TURBO toggle: flip it off to bypass the LoRA switch entirely and fall back to the full ~20-step base ref2va model, regardless of which LoRA is selected in the dropdown.
- Everything (model loading, CLIP/VAE loading, sampling, turbo logic) is packed into one collapsed subgraph with a small set of exposed widgets, so the graph stays readable at a glance instead of a wall of wired nodes.
Requirements
- ComfyUI with core MiniMax H3 support comfy-core) — no custom node packs
- minimax_h3_ref2va_pruned_int8_convrot.safetensors (diffusion model)
- qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors (text encoder)
- minimax_h3_video_vae_fp16.safetensors + minimax_h3_audio_vae_fp32.safetensors (VAEs)
- Optional: one or more MiniMax H3 turbo LoRAs (filename must contain 4step or 8step for the auto-switcher to detect it — anything else falls back to 20-step base behavior even with TURBO on)
Usage
1. Load the workflow into ComfyUI.
2. In Load Image, upload your reference/subject image.
3. In Load Audio (splash SFX reference), upload your reference audio clip.
4. Set your Resolution Selector aspect ratio and megapixels.
5. Write your scene prompt describing the shot(s) and timeline — reference your inputs as <Picture 1> and <Audio 1> in the text (see the pinned note node for the subject/retention tag format this workflow expects).
6. Pick a turbo LoRA from the dropdown (or leave TURBO off to skip it) and set Duration/Frame Rate as usual.
7. Queue the prompt.
Notes / tuning
- If your output looks undercooked or noisy at low step counts, check the console for lora key not loaded warnings — that confirms whether the selected turbo LoRA actually applied its weights rather than silently mismatching against the ref2va checkpoint.
- The audio-retention instruction is appended automatically; if you want H3 to treat your reference audio more loosely (e.g. as a rhythm/atmosphere cue rather than a literal final track), edit or remove the Audio Guidance primitive string node rather than fighting it in your own prompt text.
- TURBO off is the safest fallback if a LoRA misbehaves or you're troubleshooting quality — it forces the full base step count independent of whatever's in the LoRA dropdown.
- This is ref2va, not fl2va: there's no first_framelast_frame keyframe conditioning here, only reference image(s) + reference audio. If you need locked start/end frames, that's a different H3 task type.
- H3 generations are capped at roughly 15s natively (17-frame-per-block grid at 24fps) — this workflow doesn't chain multiple generations for longer output.
Credit
Built on Comfy-Org's official MiniMax H3 ref2va template MiniMaxH3ReferenceToVideo), with the turbo-LoRA step-matching logic and audio-retention prompt injection added on top using stock ComfyUI logic/string nodes.