MiniMax H3 for ComfyUI

Required Setup

1. Install the custom node package

This workflow requires the ComfyUI-MiniMaxH3 custom node package, published on the Comfy Registry with package ID minimax-h3.

Easiest install: open ComfyUI Manager, search for ComfyUI-MiniMaxH3, install it, then restart ComfyUI.

Manual install:

git clone https://github.com/xiaolibai-sys/ComfyUI-MiniMaxH3.git ComfyUI/custom_nodes/ComfyUI-MiniMaxH3
pip install -r ComfyUI/custom_nodes/ComfyUI-MiniMaxH3/requirements.txt

Then restart ComfyUI.

2. Set API Keys

API keys are read only from environment variables. Do not paste API keys into the workflow JSON.

• MiniMax H3 Context IR Refiner requires the environment variable: IR_KEY

• MiniMax H3 OpenAI-Compatible Refiner reads the environment variable name selected in the node, for example Api_DeepSeek or Api_Kimi

After setting the environment variables, restart ComfyUI so the new keys are available to the nodes.

Core Highlights

Structured Storyboard

• Build multi-shot video plans with per-shot duration, visual prompt, camera movement, dialogue, sound, and music.

• Define global subjects once and reuse their names in later shots. The backend converts names into standard <Subject N> labels, while text inside <d>...</d> is protected from replacement.

• Dialogue speaker IDs such as (S1) and (S2) can be written manually or generated automatically by the connected refiner.

• The storyboard maps naturally to MiniMax H3 official prompt fields: subject_definitions, summary, retention_analysis, detailed_description, overall_soundscape, and non_diegetic_music.

Prompt Refiners

• Includes official MiniMax H3 Context IR Refiner support.

• Includes an OpenAI-compatible Refiner, so services such as DeepSeek or Kimi can polish storyboards without multimodal model support.

• PackageData can provide image, video, and audio references, plus textual Notes that help non-multimodal LLMs understand the media.

• A built-in preview shows the refined structured prompt before it reaches Conditioning.

AdaLN Cache

• Optional pre-bake of AdaLN modulations before sampling.

• Can unload a large portion of the AdaLN-related branch during iterative sampling, reducing peak memory pressure.

• The included workflow estimates roughly 11 GB lower occupancy with INT8 weights and around 23 GB with BF16 weights, at the cost of a one-time pre-bake of about 30 seconds.

• Designed to work alongside BlockSwap for large-model sampling on limited VRAM.

Low-VRAM Sampling

• Streaming model loading and BlockSwap with a CPU home pool and optional disk prefetch.

• Supports bf16, fp16, fp8, int8, nvfp4, and convrot checkpoint formats.

• Built-in TeaCache arguments, attention backend selection, and sampler stats.

• MiniMax H3 Unload All releases cached models, VAEs, and encoders when needed.

Included Workflow

The provided workflow is a two-shot T2VA example: a cat jumps from a sofa to a windowsill, then the camera moves outside to a courtyard where autumn leaves fall past the window.

It includes:

• Model and weight placement notes

• A structured Storyboard

• An OpenAI-compatible Refiner chain

• Joint video/audio Conditioning

• BlockSwap and AdaLN sampling settings

• Video Helper Suite output

Requirements

• ComfyUI

• MiniMax H3 model weights, text encoder, and video/audio VAE files

• API key only when using the Context IR or OpenAI-compatible Refiner nodes

Tags

ComfyUI, MiniMax H3, T2VA, I2VA, FL2VA, L2VA, Ref2VA, Storyboard, Video Generation, Audio Generation