MiniMax H3 Ref2VA — Audio-Driven Turbo (Image + Audio to Video)
A ComfyUI workflow for MiniMax H3 that uses your own reference audio instead of the model’s invented sound. It’s built on H3’s ref2va (reference-to-video/audio) task, the only H3 task whose node accepts an audio reference. You can feed it dialogue for lipsync, an SFX cue, ambience or music, along with one or two reference images, and the audio is built into the generated clip’s timing.
It uses only stock ComfyUI core nodes. No custom node packs are needed.
WHAT’S NEW IN V2
Audio Start (s): begin the reference audio at any point in the song or clip. It’s trimmed automatically to the video length. The audio must be at least start + duration seconds long.
Second reference image: use
<Picture 1>and<Picture 2>in your prompt. Leave the second loader empty or bypassed if you only need one.Default duration is now 5 s, which makes testing quicker.
Clearer node labels, and the pinned note in the workflow documents the new options.
WHAT IT DOES
Takes one or two reference images, a reference audio clip and a text prompt, and generates a video conditioned on all of them through MiniMaxH3ReferenceToVideo.
Automatically appends an audio-retention instruction to your prompt (
<Audio 1>: fully_copy, meaning<Audio 1>is reused 1:1 as the video’s complete final audio track). The model is told to keep your audio as the literal final track, not treat it as a loose style cue, and you don’t have to type this yourself.Turbo LoRA switcher: pick any turbo LoRA from the dropdown and the workflow reads the filename for “8step” or “4step” and sets the matching step count automatically, so steps can’t drift out of sync with the LoRA you chose.
Turbo toggle: turn it off to skip the LoRA entirely and run the full ~20-step base model, whichever LoRA is selected.
Everything (model loading, CLIP/VAE loading, sampling, turbo logic, audio trimming) is packed into one subgraph with a small set of exposed settings, so the graph stays readable.
REQUIREMENTS
ComfyUI with core MiniMax H3 support (comfy-core), no custom nodes
minimax_h3_ref2va_pruned_int8_convrot.safetensors (diffusion model)
qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors (text encoder)
minimax_h3_video_vae_fp16.safetensors and minimax_h3_audio_vae_fp32.safetensors (VAEs)
Optional: one or more MiniMax H3 turbo LoRAs. The filename must contain “4step” or “8step” for the auto-switcher to detect it. Anything else falls back to the 20-step base behavior, even with turbo on.
HOW TO USE
Load the workflow into ComfyUI.
In Load Reference Image 1, upload your reference image. Optionally add a second in Load Reference Image 2.
In Load Audio (SFX or voice reference), upload your audio clip.
Set Audio Start (s) if you want to begin partway into the audio.
Choose your aspect ratio and megapixels in the Resolution Selector.
Write your scene prompt describing the shots and timeline. Refer to your inputs as
<Picture 1>,<Picture 2>and<Audio 1>. The pinned note in the workflow shows the subject/retention tag format it expects.Pick a turbo LoRA from the dropdown (or turn turbo off), and set duration and frame rate.
Queue the prompt.
NOTES AND TUNING
If the output looks undercooked or noisy at low step counts, check the console for “lora key not loaded” warnings. They show whether the turbo LoRA actually applied or mismatched the ref2va checkpoint.
Turbo off is the safest fallback if a LoRA misbehaves or you’re troubleshooting quality. It forces the full base step count.
The audio-retention instruction is added automatically. To make H3 treat your audio more loosely (as a rhythm or atmosphere cue instead of the literal final track), edit or remove the Audio Guidance string node rather than fighting it in your prompt.
This is ref2va, not first-frame/last-frame. There’s no keyframe conditioning, only reference image(s) and reference audio. For locked start and end frames you need a different H3 task type.
H3 generations are capped at roughly 15 s natively (17-frame blocks at 24 fps). 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, audio-retention prompt injection, audio start time and second reference image added using stock ComfyUI nodes.
Description
V2 changes
Audio Start (s): new setting on the main node. It starts the reference audio at any point in the song or clip instead of always at 0:00. The audio is automatically trimmed to the video length. The clip must be at least start + duration seconds long.
Second reference image: there’s now a second image loader, so you can use
<Picture 1>and<Picture 2>in your prompt. Leave the second one empty or bypassed if you only need one.Default duration: now 5 s, which is easier to test with. Raise it once you know what your GPU can handle.
Labels: the loaders are renamed (“Load Reference Image 1”, “Load Reference Image 2”, “Load Audio (SFX or voice reference)”), and the note in the workflow documents the new options.