<h1 id="minimax-h3-for-comfyui">MiniMax H3 for ComfyUI</h1><h2 id="required-setup">Required Setup</h2><h3 id="1.-install-the-custom-node-package">1. Install the custom node package</h3><p>This workflow requires the <strong>ComfyUI-MiniMaxH3</strong> custom node package, published on the Comfy Registry with package ID <code>minimax-h3</code>.</p><p>Easiest install: open ComfyUI Manager, search for <code>ComfyUI-MiniMaxH3</code>, install it, then restart ComfyUI.</p><p>Manual install:</p><pre><code>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</code></pre><p>Then restart ComfyUI.</p><h3 id="2.-set-api-keys">2. Set API Keys</h3><p>API keys are read only from environment variables. Do not paste API keys into the workflow JSON.</p><ul><li><p><strong>MiniMax H3 Context IR Refiner</strong> requires the environment variable: <code>IR_KEY</code></p></li><li><p><strong>MiniMax H3 OpenAI-Compatible Refiner</strong> reads the environment variable name selected in the node, for example <code>Api_DeepSeek</code> or <code>Api_Kimi</code></p></li></ul><p>After setting the environment variables, restart ComfyUI so the new keys are available to the nodes.</p><h2 id="core-highlights">Core Highlights</h2><h3 id="structured-storyboard">Structured Storyboard</h3><ul><li><p>Build multi-shot video plans with per-shot duration, visual prompt, camera movement, dialogue, sound, and music.</p></li><li><p>Define global subjects once and reuse their names in later shots. The backend converts names into standard <code>&lt;Subject N&gt;</code> labels, while text inside <code>&lt;d&gt;...&lt;/d&gt;</code> is protected from replacement.</p></li><li><p>Dialogue speaker IDs such as <code>(S1)</code> and <code>(S2)</code> can be written manually or generated automatically by the connected refiner.</p></li><li><p>The storyboard maps naturally to MiniMax H3 official prompt fields: <code>subject_definitions</code>, <code>summary</code>, <code>retention_analysis</code>, <code>detailed_description</code>, <code>overall_soundscape</code>, and <code>non_diegetic_music</code>.</p></li></ul><h3 id="prompt-refiners">Prompt Refiners</h3><ul><li><p>Includes official MiniMax H3 Context IR Refiner support.</p></li><li><p>Includes an OpenAI-compatible Refiner, so services such as DeepSeek or Kimi can polish storyboards without multimodal model support.</p></li><li><p>PackageData can provide image, video, and audio references, plus textual Notes that help non-multimodal LLMs understand the media.</p></li><li><p>A built-in preview shows the refined structured prompt before it reaches Conditioning.</p></li></ul><h3 id="adaln-cache">AdaLN Cache</h3><ul><li><p>Optional pre-bake of AdaLN modulations before sampling.</p></li><li><p>Can unload a large portion of the AdaLN-related branch during iterative sampling, reducing peak memory pressure.</p></li><li><p>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.</p></li><li><p>Designed to work alongside BlockSwap for large-model sampling on limited VRAM.</p></li></ul><h3 id="low-vram-sampling">Low-VRAM Sampling</h3><ul><li><p>Streaming model loading and BlockSwap with a CPU home pool and optional disk prefetch.</p></li><li><p>Supports bf16, fp16, fp8, int8, nvfp4, and convrot checkpoint formats.</p></li><li><p>Built-in TeaCache arguments, attention backend selection, and sampler stats.</p></li><li><p><code>MiniMax H3 Unload All</code> releases cached models, VAEs, and encoders when needed.</p></li></ul><h2 id="included-workflow">Included Workflow</h2><p>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.</p><p>It includes:</p><ul><li><p>Model and weight placement notes</p></li><li><p>A structured Storyboard</p></li><li><p>An OpenAI-compatible Refiner chain</p></li><li><p>Joint video/audio Conditioning</p></li><li><p>BlockSwap and AdaLN sampling settings</p></li><li><p>Video Helper Suite output</p></li></ul><h2 id="requirements">Requirements</h2><ul><li><p>ComfyUI</p></li><li><p>MiniMax H3 model weights, text encoder, and video/audio VAE files</p></li><li><p>API key only when using the Context IR or OpenAI-compatible Refiner nodes</p></li></ul><h2 id="tags">Tags</h2><p>ComfyUI, MiniMax H3, T2VA, I2VA, FL2VA, L2VA, Ref2VA, Storyboard, Video Generation, Audio Generation</p>

Files with SHA256: e3e80a05278e8cd84fb21d911968284560eef7bfc12efb32f74136b2718b42e2