Anyone who has ever built a video model workflow in ComfyUI knows how it ends: twenty nodes, fifteen wires, three custom packs — and one update is enough to blow the whole canvas apart. With MiniMax H3, which generates video complete with audio, the graph gets even wilder — separate branches for the text encoder, the video VAE, the audio VAE, Motion Context… That's why I got excited when I stumbled upon the ComfyUI-ALLinONE-MinimaxH3 repository. It packs the whole circus into a single node.
What MiniMax H3 is and why it deserves attention
MiniMax H3 is the current generation of the Chinese video model that generates video including native audio — speech, sound effects and music in a single pass, no separate dubbing step. The weights are freely available via Comfy-Org in pruned, quantized versions (int8 for the diffusion models, nvfp4 for the text encoder), so it runs locally on a gaming GPU. No API key, no credits, no sending your data to someone else's server.
The model comes in two variants: fl2va for classic text-to-video and frame animation, and ref2va for reference-driven generation — from images, videos or audio.
Eight modes in one node
After installation, just double-click the canvas, search for "ALL in ONE MiniMaxH3" and you get a single node that switches between eight modes:
- Image — text-to-image, image editing or mixing up to 9 references via H3 Studio
- T2V — text to video with native audio
- I2V — animating a still image, optionally morphing into an end frame
- R2V — reference-driven generation (images, videos and audio)
- Audio Drive — video driven by an audio track, including lip-sync
- Keyframes — your own stills placed at arbitrary positions on the timeline
- Extend — seamless continuation of an existing video
- Chain — chaining multiple clips via H3 Motion Context; the handoff happens in latent space, so nothing gets re-encoded and quality doesn't degrade between clips
On top of that, the author bundled things you would normally solve with extra packs in ComfyUI: live preview during sampling, a searchable history that lets you reuse old prompts, and an output library with favorites and RTX upscale integration.
Installation and what you'll need
The node itself installs the usual way:
cd ComfyUI/custom_nodes
git clone https://github.com/LeonQ8/ComfyUI-ALLinONE-MinimaxH3.gitThen restart ComfyUI and the node shows up. It won't run without models, though — the base for T2V, I2V and R2V is five files from the Comfy-Org/MiniMax-H3 repository:
minimax_h3_fl2va_pruned_int8_convrot.safetensors→diffusion_models/minimax_h3_ref2va_pruned_int8_convrot.safetensors→diffusion_models/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors→text_encoders/minimax_h3_video_vae_fp16.safetensorsandminimax_h3_audio_vae_fp32.safetensors→vae/
A practical detail I appreciate: additional node packs are only needed for the modes you actually use. Audio Drive wants comfyui-vrgamedevgirl, the Keyframes/Extend/Chain modes need ComfyUI-H3-Motion-Context-MultiRef, and Image mode needs ComfyUI-MiniMax-H3-Studio. If you want faster generation, grab the Turbo LoRA minimax_h3_turbo_v4_step600_ema.safetensors and the matching preset pack. None of this is required for basic text-to-video.
What to watch out for
The project is officially in beta and the author develops against a precisely pinned stack: ComfyUI 0.33.0, Python 3.12.10, PyTorch 2.9.1 with CUDA 13.0, Windows portable. That doesn't mean it won't run elsewhere, but when something breaks, the first step is to compare your versions against his COMPATIBILITY.md — he says himself that after an update it's almost always ComfyUI or one of the node packs that moved, not this node. From my own experience with ComfyUI, I can only co-sign that.
Two specific known landmines:
- The ComfyUI-MiniMaxH3-Cache accelerator pack breaks the Chain and Extend modes on ComfyUI 0.33 because it changes model signatures. The fix is simple — rename the pack's folder to
.disabledwhen you need those modes. - ComfyUI builds from August 6–13, 2026 contained a bug that crashed R2V mode with a tensor shape mismatch. Fixed in core on August 13, so just update ComfyUI.
The author doesn't state VRAM requirements. The models are pruned and quantized, but the text encoder is derived from a 32B Qwen — I wouldn't expect miracles on an 8GB card; this is a toy for 16 GB and up.
The license is GPL-3.0, so the node is free even for commercial video work; only if you wanted to build on it and distribute your own software would that software have to be open-source too.
Why I think it makes sense
There have been a few "all-in-one" attempts in ComfyUI before, and most of them failed by hiding too many settings. This node takes a smarter approach — it keeps the parameters accessible and only removes the manual wiring between components that everyone connected the same way anyway. And crucially, it also covers advanced disciplines like latent-space clip chaining, which a regular user would never assemble from the official workflows on their own.
If you're just starting with MiniMax H3, begin with T2V mode and the five base models, and only then add packs for the other modes — you'll save yourself the hunt for which of seven freshly installed add-ons broke your queue. And before your next ComfyUI update, check COMPATIBILITY.md in the repository to see whether a known trap is waiting for you.
Repository link: github.com/LeonQ8/ComfyUI-ALLinONE-MinimaxH3 (GPL-3.0, free)
