MiniMax H3 ComfyUI Workflow: Files, Audio, and 2K Limits

The model you can run locally in ComfyUI is MiniMax H3, not H3 Max. Start with ComfyUI 0.30.0 or newer, load an official H3 template, and install four parts: an FL2VA or Ref2VA diffusion model, the Qwen3-VL text encoder, the video VAE, and the audio VAE. Leave any one of them out and the graph is either incomplete or unable to produce the intended picture-and-sound file.

That naming distinction matters. MiniMax H3 is the open-weight base model. H3 Max and H3 Max Turbo are separate hosted routes. A tutorial that calls a downloadable checkpoint “H3 Max” is collapsing two different products into one label.

Choose the workflow before downloading 40 GB of files

There are three sensible ways to use this model family. Pick the operational model first; the interface can come second.

The local and browser paths are not two front ends for the same checkpoint. The H3 versus H3 Max guide compares their output contracts in detail.

Use FL2VA for a first run; add Ref2VA only when the job needs it

MiniMax publishes two H3-Base task families:

If your first goal is “make this still image move,” FL2VA is the smaller decision. Ref2VA is not an upgrade switch for every shot; it is a different checkpoint for a different input contract. Download it when a character, motion clip, camera move, voice, or other reference is part of the brief.

ComfyUI's current MiniMax H3 guide lists five templates in Template Library → Video: text-to-video, image-to-video, reference-to-video, multiframe reference, and Fun ControlNet Union. The first three need ComfyUI 0.30.0 or newer. Multiframe Reference needs 0.34.0+, while Fun ControlNet Union starts at 0.35.0+ and Union 2.0 needs 0.38.0+.

For a clean first run, choose the official text-to-video template. Do not begin by combining a quant, a speed LoRA, ControlNet, a custom sampler, and three references. A small known-good graph gives every later change something to compare against.

Put each model file in the directory its loader scans

The template's download dialog is the safest route because the repository changes over time. If you install manually, the current Comfy-Org H3 model card maps a practical FL2VA stack like this:

Checked October 29, 2026, those four public files total about 40.1 decimal GB (37.3 GiB) of downloads. Adding the matching pruned INT8 Ref2VA diffusion checkpoint adds about 21.0 decimal GB. These are file sizes, not a promise that the same amount of VRAM is sufficient for every duration, canvas, sampler, and offload setting.

The current model card prefers int8_convrot diffusion files when a PyTorch cu130 build is available. It says to use fp8_scaled when that path is not available. Treat the filename as a compatibility choice, not a quality badge: a graph cannot compensate for a quantization format that the installed runtime cannot load.

After copying files, restart ComfyUI or refresh its model list. Then open every model dropdown and select the exact file once. A file sitting in models/ but not in the loader's specific subdirectory is effectively invisible.

Use a prompt that reveals whether both halves of H3 work

A first render should make motion and sound easy to inspect. This short diagnostic brief keeps the camera simple and gives the audio decoder unmistakable events:

integrated_multimodal_description: [Shot 1] Locked medium-wide shot inside a quiet clockmaker's workshop at blue hour. A small brass wind-up bird stands on a wooden bench. At 00:01 its key turns three times, its wings open, and it hops once toward the camera. Keep the bird, bench, framing, and cool window light consistent for the whole shot.

overall_soundscape: Three crisp mechanical clicks from the winding key, a soft metal wing flutter, one wooden tap when the bird lands, and faint rain against the window. No speech.

non_diegetic_music: N/A.

Use the template's ordinary settings for this smoke test. If the framing moves, the bird changes shape, or one sound event is absent, you have a useful baseline to debug. Changing resolution, sampler, duration, model, prompt, and acceleration at the same time destroys that baseline.

Fix the failure by locating the missing stage

The H3 templates or nodes are missing

Check the ComfyUI version first. Text-to-video, image-to-video, and reference-to-video require 0.30.0 or newer. Update ComfyUI, restart it, then look under Template Library → Video before installing an unrelated custom-node pack.

A model dropdown is red or empty

Match the file to the loader directory. The diffusion checkpoint belongs in diffusion_models, the Qwen file in text_encoders, and both VAEs in vae. Restart after moving them. Also confirm that the workflow expects FL2VA rather than Ref2VA.

The MP4 has picture but no H3-generated sound

Confirm that minimax_h3_audio_vae_fp32.safetensors is installed, selected by the audio decoder, and still connected in the graph. The video VAE cannot replace it. Check the exported MP4 in a player with its own volume control as well; a muted browser preview is not evidence that the audio track is missing.

The reference image, video, or voice has little effect

FL2VA is for text and keyframes. Use the reference-to-video template and Ref2VA checkpoint when the input is supposed to control identity, motion, camera behavior, or voice. In the prompt, say what each asset controls instead of attaching files without roles.

The graph runs out of memory

Return to one official template and one compatible checkpoint. Reduce the canvas or duration, remove optional branches, and let ComfyUI offload before adding acceleration nodes. The official ComfyUI work made H3 runnable on much smaller hardware than the full-precision model, but it does not make every graph fit every GPU.

There is no local 2K setting

That is expected in the current official workflow. ComfyUI's latest guide describes the open weights as targeting roughly one megapixel—about a 768-pixel short edge—at 24 fps. MiniMax's official model card describes its full 2K workflow as a three-stage system: hosted Context-IR, a local H3-Base pass, and hosted H3 Regenerate-2K.

In other words, “H3 supports 2K” does not mean a local FL2VA graph has a native 2K dropdown. The H3 Regenerate 2K guide explains the eligible source file, final-prompt manifest, input rebilling, and separate regeneration request.

Check the territory and revenue terms before commercial use

Open weights do not mean unrestricted commercial use. MiniMax's current H3 Community License grants a limited, royalty-free license inside its Applicable Territory when its conditions are met. That territory excludes the United States, European Union, United Kingdom, and South Korea. Commercial products and services above $20 million in yearly revenue also need separate prior authorization from MiniMax.

Comfy says users in the excluded territories can obtain a commercial license through its program, while Comfy Cloud generations already include commercial rights. Check the live MiniMax license and the terms of the path you actually use before shipping client work or a paid product.

This is another reason to separate local H3, Comfy Cloud, and hosted H3 Max. They can share a node editor or a family name while having different weights, infrastructure, pricing, and license paths.

The short answers

Is H3 Max available as a ComfyUI checkpoint?

The official downloadable ComfyUI weights are MiniMax H3-Base FL2VA and Ref2VA. H3 Max and H3 Max Turbo are separate hosted model routes, not names for those local files.

Which H3 workflow should I install first?

Use the official text-to-video or image-to-video FL2VA template. Add Ref2VA only when reference images, video, or audio are central to the job.

Why is my MiniMax H3 ComfyUI video silent?

First check the audio VAE file, its selected loader, and the graph connection into audio decoding. Then inspect the exported MP4 outside a muted browser preview.

Can local ComfyUI H3 generate 2K?

The current official local target is roughly 1 MP/768-short-edge. MiniMax's documented full 2K process combines local H3-Base with hosted Context-IR and Regenerate-2K stages; it is not a single local resolution switch.

Start with one graph you understand. A working four-file FL2VA pipeline is more valuable than a crowded canvas whose model, prompt, audio, and memory failures cannot be separated.