MiniMax H3 ComfyUI Guide
MiniMax H3 has native ComfyUI workflows for text, keyframes, and multimodal references. Use this guide to understand the local model stack, or start generating here without downloading weights, configuring CUDA, or using local VRAM.
MiniMax H3 ComfyUI / verified August 7, 2026
From an empty graph to the right H3 workflow
The MiniMax H3 ComfyUI workflow is a native set of templates and nodes for text-to-video, image-to-video, and reference-to-video generation with synchronized stereo audio. Local use gives you graph-level control, but it also requires current ComfyUI, several large model files, compatible GPU software, storage, and memory planning. This browser workspace offers the same three input decisions as a managed generation path, not as a ComfyUI replacement or download mirror.
Start with the official native templates
Update ComfyUI to version 0.30.0 or later, open Template Library → Video, and choose a MiniMax H3 T2V, I2V, or R2V example. The templates are the safest starting point because they reveal the current native node contract and prompt you for the matching model files. A third-party custom node may add convenience, but it is not required merely to access the official H3 workflows.
Treat a downloaded workflow JSON as executable configuration, not as a picture of colored boxes. Open it only after updating ComfyUI, inspect missing-node warnings, verify every model filename, and confirm the output path before queueing. If a community graph was built for GGUF, FP8, INT8, a custom loader, or an older nightly build, do not assume that its node types and memory behavior match the official template.
Place four model roles in the correct folders
The native T2V and I2V examples use the pruned INT8 ConvRot FL2VA diffusion model, while R2V uses a separate Ref2VA diffusion model. Diffusion weights belong in `ComfyUI/models/diffusion_models/`. The shared Qwen3-VL text encoder belongs in `models/text_encoders/`; the video FP16 VAE and audio FP32 VAE both belong in `models/vae/`. A filename in the wrong directory looks like a missing-model problem even when the download is complete.
FL2VA covers an open text prompt plus zero, one, or two keyframes. Ref2VA is a different weight path built for images, videos, and audio references. Downloading only the FL2VA stack will not make an R2V graph complete. Keep a manifest of exact filenames and sources, and avoid casually renaming safetensors: loader dropdowns, workflow metadata, and troubleshooting guides often rely on those names.
Choose local control or a managed browser run
A local graph is valuable when you need node-level experimentation, offline assets, quantized weights, custom attention patches, or repeatable automation on hardware you control. The cost is operational: model downloads, disk capacity, CUDA and PyTorch compatibility, VRAM offload, system RAM, startup time, and maintenance when nodes change. Native support simplifies the graph; it does not make a 33B video model lightweight.
The generator on this page is hosted. It requires no local GPU or ComfyUI install and exposes text, first/last-frame, and multimodal-reference paths through a focused form. It does not download JSON workflows, reveal a local node graph, or load your custom nodes. Use it to create and compare H3 shots quickly; choose local ComfyUI when graph ownership is the actual requirement.
Set resolution, duration, references, and sound deliberately
The official Resolution Selector computes width and height from aspect ratio, megapixels, and a multiple of 32. Around 1.0 megapixel at 16:9 produces roughly 1344×768, H3’s native 768-pixel short-edge canvas. Higher dimensions and longer clips increase memory and runtime. Duration follows the model’s frame grid at 24 FPS, so a UI may snap a requested value to a compatible frame count.
For R2V, label references in connection order—such as `<Picture 1>`, `<Video 1>`, and `<Audio 1>`—and give each one a single job: identity, wardrobe, camera rhythm, motion, ambience, or voice. The published limits are nine images, three videos, and three audio clips, with twelve files total. More inputs can create conflicts; a smaller, clearly assigned reference set is usually easier to debug.
Which ComfyUI path fits your goal?
Native template learner. Start with the official graph, change one parameter at a time, and save a known-good JSON before adding custom loaders or attention patches.
Local pipeline builder. Use ComfyUI when deterministic graph reuse, private local assets, quantization tests, or downstream nodes matter more than setup time.
Browser-first creator. Use the hosted generator when the goal is an H3 result rather than owning the graph, model files, drivers, and memory plan.
Continue through the MiniMax H3 technical guides
Know what every file does before loading it.
The native stack separates diffusion weights, a Qwen3-VL text encoder, video decoding, and audio decoding. Choose FL2VA or Ref2VA before downloading the largest file.
Update, load, then test small.
Use ComfyUI 0.30.0 or later and open the native H3 template before installing optional community nodes. This gives you a known-good graph contract.
Put diffusion, text-encoder, video-VAE, and audio-VAE files in their documented folders. Confirm the loader dropdowns resolve exact filenames.
Run a short preview at a conservative resolution. Verify video, stereo audio, and output path before increasing duration, megapixels, or reference complexity.
Want the H3 result without the local stack?
Use the generator above for text, keyframe, or multimodal-reference video without downloading weights or configuring a GPU.
Hosted H3 plans instead of local hardware setup
Use credits for managed generation when maintaining models, drivers, disk, and GPU memory is not part of the job.
Standard
Billed $238.80 yearly
- Commercial Usage RightsYearly Only
- Up to 624 videos
- Video models low to $0.11/s
- Up to 9,360 images
- All AI Models
- Priority Support
- Priority Processing Speed
- Remove Watermark
- Free Video UpscalerYearly Only
- Free Frame InterpolationYearly Only
- More Free AI ToolsYearly Only
Business
Billed $418.80 yearly
- Commercial Usage RightsYearly Only
- Up to 1,360 videos
- Video models low to $0.09/s
- Up to 20,400 images
- All AI Models
- Priority Support
- Dedicated Support
- Priority Processing Speed
- Remove Watermark
- Free Video UpscalerYearly Only
- Free Frame InterpolationYearly Only
- More Free AI ToolsYearly Only
Enterprise
Billed $898.80 yearly
- Commercial Usage RightsYearly Only
- Up to 3,920 videos
- Video models low to $0.06/s
- Up to 58,800 images
- All AI Models
- Priority Support
- Dedicated Support
- Priority Processing Speed
- Remove Watermark
- Free Video UpscalerYearly Only
- Free Frame InterpolationYearly Only
- More Free AI ToolsYearly Only
Included models
Ideogram 4
Z-Image Turbo
Grok Imagine Image
Nano BananaNano Banana ProNano Banana 2
GPT Image 2
Seedream 5.0 LiteSeedream 4.5
MiniMax H3Seedance 2.0Seedance 2.0 FastSeedance 2.0 MiniSeedance 2.5Coming SoonGet Early BirdSeedance 1.5 ProGrok Imagine Video
Gemini Omni FlashVeo 3.1 QualityVeo 3.1 LiteVeo 3.1 FastWan 2.7Wan 2.6Wan 2.5Credits activate instantly for MiniMax H3 generation workflows.
A workflow is healthy when every dependency is explainable.
Record ComfyUI version, workflow source, model filenames, loader types, resolution, duration, and attention backend before troubleshooting output quality.
Try the Browser GeneratorMiniMax H3 ComfyUI FAQ
What is the MiniMax H3 ComfyUI workflow?
It is ComfyUI’s native node-and-template path for MiniMax H3 text-to-video, image-to-video, and reference-to-video generation with stereo audio. The official template library provides starting graphs, while FL2VA and Ref2VA use different diffusion weights for different input modes.
Which files does MiniMax H3 need in ComfyUI?
A native workflow needs an FL2VA or Ref2VA diffusion model, the published Qwen3-VL H3 text encoder, a video VAE, and an audio VAE. Diffusion weights, text encoders, and VAEs go in their corresponding folders under `ComfyUI/models/`.
Does MiniMax H3 require custom ComfyUI nodes?
No custom node is required to start with the official native H3 templates in a current ComfyUI release. Optional community loaders, quantization formats, Sage Attention patches, and workflow utilities may add features, but they also add version and compatibility variables.
Can I run MiniMax H3 without a local GPU?
Yes—by using a hosted service rather than running the open weights locally. The generator on this page uses managed infrastructure, so your browser does not need CUDA or local VRAM. It does not provide a local ComfyUI graph or model downloads.
Why does an imported H3 workflow show missing models?
The usual causes are an outdated ComfyUI build, a diffusion weight for the wrong H3 mode, files placed in the wrong model folder, renamed safetensors, or community nodes that are not installed. Resolve the official template first, then add optional components one at a time.