<div data-youtube-video><iframe width="640" height="480" allowfullscreen="true" autoplay="false" disablekbcontrols="false" enableiframeapi="false" endtime="0" ivloadpolicy="0" loop="false" modestbranding="false" origin playlist src="https://www.youtube.com/embed/UBgrzyj1sfg?rel=1" start="0"></iframe></div>Full video tutorial: every setting explained, plus the 12 GB VRAM setup step by step: <a target="_blank" rel="ugc" href="https://youtu.be/UBgrzyj1sfg">https://youtu.be/UBgrzyj1sfg</a>The fastest MiniMax H3 setup I've found: the lightx2v 4-step turbo LoRA + the FastVideo sparse attention (VSA) gate + the Semantic Bridge adapter, text to video and image to video with native audio in one graph.Measured on my test GPU (RTX A4500 20 GB, PyTorch cu130):<ul><li>Text to video, 5 s, 640x960 (0.6 MP): 0:55 sampling, peak about 12.8 GB VRAM (a 12 GB limit still ran at full speed, 0:57)</li><li>Image to video, 1 MP: 2:08 sampling, 12.2 GB peak (on 12 GB cards use 0.6 MP: 1:04, 11.3 GB peak)</li><li>8 GB: not recommended (the sampler needs about 9 GB and spills into system RAM)</li></ul>Settings that work for BOTH modes: sampler er_sde + scheduler sgm_uniform, 4 steps, MiniMax Sigma Shift ON (6 / 3). Already set in this file.Top 3 tips<ol><li>Use PyTorch cu130 (driver 580+). On cu128 ComfyUI disables its fast kernels: 4:30 vs 1:37 sampling, same workflow.</li><li>A full Task Manager is normal: ComfyUI fills free VRAM on purpose.</li><li>For image to video, use a full-size image and match the aspect ratio.</li></ol>Models (Hugging Face):<ul><li>MiniMax H3 by MiniMax: <a target="_blank" rel="ugc" href="https://huggingface.co/MiniMaxAI/MiniMax-H3">https://huggingface.co/MiniMaxAI/MiniMax-H3</a></li><li>ComfyUI files (ref2va pruned int8 or w6a8, Qwen3-VL text encoder, video + audio VAE) by Comfy-Org: <a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/MiniMax-H3/tree/main">https://huggingface.co/Comfy-Org/MiniMax-H3/tree/main</a></li><li>4-step turbo LoRA by lightx2v: <a target="_blank" rel="ugc" href="https://huggingface.co/lightx2v/Minimax-h3-Turbo/tree/main">https://huggingface.co/lightx2v/Minimax-h3-Turbo/tree/main</a></li><li>FastVideo VSA gate file: <a target="_blank" rel="ugc" href="https://huggingface.co/barelymining/ComfyUI-MiniMax-H3-FastVideo/tree/main">https://huggingface.co/barelymining/ComfyUI-MiniMax-H3-FastVideo/tree/main</a></li><li>Semantic Bridge adapter + node (zip, not in the Manager) by speach1sdef178: <a target="_blank" rel="ugc" href="https://huggingface.co/speach1sdef178/MiniMax-H3-Semantic-Bridge/tree/main">https://huggingface.co/speach1sdef178/MiniMax-H3-Semantic-Bridge/tree/main</a></li></ul>Custom nodes: KJNodes, rgthree, Video Helper Suite, Memory Cleanup, Nvidia RTX Nodes (ComfyUI Manager &gt; Install Missing Custom Nodes), plus:<ul><li>Ref2VA VSA node by Kablex: <a target="_blank" rel="ugc" href="https://github.com/Kablex/ComfyUI-Ref2VA-VSA">https://github.com/Kablex/ComfyUI-Ref2VA-VSA</a></li><li>Optional live VRAM monitor, Crystools by crystian: <a target="_blank" rel="ugc" href="https://github.com/crystian/ComfyUI-Crystools">https://github.com/crystian/ComfyUI-Crystools</a></li></ul>Want the one-click setup? My MiniMax H3 All-In-One Windows installer sets up ComfyUI, the custom nodes and the models for you automatically:<ul><li>Patreon (one-click installer): <a target="_blank" rel="ugc" href="https://www.patreon.com/TheLocalLab/posts/minimax-h3-all-167996909">https://www.patreon.com/TheLocalLab/posts/minimax-h3-all-167996909</a></li><li>Individual purchase (locallabdigest.com): <a target="_blank" rel="ugc" href="https://www.locallabdigest.com/product-page/minimax-h3-all-in-one-comfyui-one-click-windows-installer">https://www.locallabdigest.com/product-page/minimax-h3-all-in-one-comfyui-one-click-windows-installer</a></li></ul>Sampler comparison (community ratings for H3 4-step, by HouPeiyi): <a target="_blank" rel="ugc" href="https://darkstarrddev.us.ci/">https://darkstarrddev.us.ci/</a>Credits: MiniMax (H3), Comfy-Org (ComfyUI files), lightx2v (turbo LoRA), FastVideo team / hao-ai-lab + Kablex (VSA node), speach1sdef178 (Semantic Bridge), HouPeiyi (sampler benchmark), crystian (Crystools). Demo images generated locally with Qwen Image 2.1.License: MiniMax H3 is released under the MiniMax H3 license, which limits where the model and its outputs may be used. Read it before you use it: <a target="_blank" rel="ugc" href="https://huggingface.co/MiniMaxAI/MiniMax-H3">https://huggingface.co/MiniMaxAI/MiniMax-H3</a>More from me:<ul><li>YouTube: <a target="_blank" rel="ugc" href="https://www.youtube.com/@TheLocalLab">https://www.youtube.com/@TheLocalLab</a></li><li>Patreon: <a target="_blank" rel="ugc" href="https://www.patreon.com/TheLocalLab">https://www.patreon.com/TheLocalLab</a></li><li>X: <a target="_blank" rel="ugc" href="https://x.com/TheLocalLab_">https://x.com/TheLocalLab_</a></li></ul>

MiniMax H3 4-Step Fast Video | T2V + I2V with Audio | ~1 min, 12GB VRAM | FastVideo VSA + Semantic Bridge

First release. Sampler er_sde + scheduler sgm_uniform, 4 steps, MiniMax Sigma Shift ON (6 / 3): works for both text to video and image to video. lightx2v 4-step turbo LoRA, FastVideo VSA gate at 75 % sparsity and Semantic Bridge v1 wired in. Model links and settings are in the notes inside the graph.

Files with SHA256: 73e4779a8fb95cfa7d8f85e55e2d2290cc2c2051c851d2b0bea1122d2f91bacc