MiniMax H3 | Forge Neo Extension

Generate videos with sound with MiniMax H3 inside Forge Neo, using the checkpoint, preset and Generate button you already know. Describe the scene and its sounds, start from a picture, give it pictures, videos and sounds of the people and things to show, or make it follow the motion of another video, and get an MP4 with picture and audio made together.

Work in progress. Tested on a 48 GB A40, a 24 GB RTX 4090 and a 16 GB RTX 2000 Ada with 32 GB of RAM, using the smaller files. Cards under 16 GB have not been tested. Requires an up-to-date Forge Neo (the neo branch, revision d70373e of 3 October 2026 or later, with comfy-kitchen 0.2.37). On an older version it stays disabled and says so in the console, and Forge Neo keeps working as usual.

Features 🚀

🔊 Video and Sound Together

• Describe what happens and what it sounds like in one prompt: footsteps, rain, engines, music, a line of dialogue

• Up to 15 seconds per clip at 24 frames per second, saved as one MP4 with picture and sound

• Audio shift slider, a silent option and a still image output

🖼️ First and Last Frame

• img2img: your input image becomes the first frame of the video

• Last frame: turn on ImageStitch Integrated, which Forge Neo already has, and add one image to its gallery

🧩 Reference Pictures, Videos and Sounds (Ref2VA)

• Up to 9 pictures, 3 videos and 3 audio clips of people, places and objects, kept in a new scene

• Guide: anchor a video or a voice at any frame of the clip, and keep the original sound of a reference video

🎯 Motion Control and Inpainting

• Fun ControlNet: the clip follows the motion and shapes of any video, through pose, depth, canny, HED, lines or scribble maps made by Forge's own preprocessors

• Video inpainting: redraw only the part of a video under a mask and keep the rest

⚡ Faster Generation

• Turbo and Acc 8-Step LoRAs: 8 steps instead of 20, about twice as fast

• FastH3 distilled checkpoint, Sparse Attention and Spectrum Integrated (21 to 25% faster)

• Live preview while the clip is generated, and an Upscale of the finished video with Forge's upscalers

🧠 Forge Memory Management

• 16 GB cards with 32 GB of RAM, and 24 GB cards with Forge's own offloading

• Smaller W4A8 and INT4 files, GGUF checkpoints and a GGUF text encoder for less RAM and disk

🛡️ Safe by Design

• Changes no Forge Neo file: remove the extension and everything is as it was

• Checks your Forge Neo at startup and every model file before loading it. Other models are not affected

Installation 📦

1. Open Forge Neo WebUI and go to Extensions → Install from URL

2. Paste: https://github.com/eduardoabreu81/minimax-h3-forge-neo

3. Click Install and restart the WebUI

4. Place the H3 checkpoint, text encoder, video VAE and audio VAE in the usual folders (the full file list is in the README)

5. Make sure FFmpeg is installed, then pick the h3 UI preset, the checkpoint, and all three of the text encoder, the video VAE and the audio VAE under VAE / Text Encoder

On a 16 GB card, start Forge Neo with --cuda-malloc. Without it, the first step can run out of VRAM.

Documentation: Wiki · File list · What's new · GitHub · License: AGPL-3.0