ALL in ONE MiniMaxH3

If you've seen MiniMax H3's official ComfyUI workflow, you know the catch with an otherwise stunning video model: it's a graph of a dozen nodes, and getting the reference packing right is an afternoon of its own. ALL in ONE MiniMaxH3 is the answer to that specific complaint. One node, a mode dropdown, your prompt or references, Generate. That's it.

The name is a bit of a lie in the best way: it doesn't call any API and needs no key. All the generation runs on local weights in your own ComfyUI. The node is a shell that builds the graph for you - the frontend assembles the proper H3 sub-workflow for whichever mode you picked, queues it through ComfyUI's normal prompt path, and reports the result back on the node itself. Under the hood it's the "one node" school of ComfyUI UX that Ján's one-node-flux-2-klein made popular last spring (that launch got 676 upvotes, mostly people tired of spaghetti graphs); this is the H3 take on it, with the official Comfy-Org H3 workflows as the base graphs.

What it can do

The mode list is the whole H3 pipeline in one widget: Image (text-to-image, edit, or a reference mix of up to 9 refs via H3 Studio), T2V (text-to-video with native audio from the fl2va model), I2V (animate a start frame, optionally morph to an end frame), R2V (reference-driven via ref2va), Audio Drive (your track is the soundtrack and drives the mouth), Keyframes (pin stills at arbitrary frame positions), Extend (continue an existing clip), Chain (multi-clip continuation with H3 Motion Context), Mask (paint or describe a region, track it, replace just that area), and Upscale (RTX VSR or SeedVR2 hooks). There's also a Library, searchable History, live preview while sampling, and a Compare & Stitch page that exports 2–4 clips side-by-side at deterministic 24 fps.

The backend node itself is almost comically small. Its only input is an optional prompt (STRING), and it returns two outputs: poster (IMAGE - the first frame as a thumbnail) and video (STRING - the relative path to the saved file). Everything else you interact with lives in the node's UI.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/LeonQ8/ComfyUI-ALLinONE-MinimaxH3.git

Restart ComfyUI, then double-click the canvas and search for ALL in ONE MiniMaxH3. ComfyUI Manager finds it under the same pack title if you prefer that route. T2V, I2V and R2V run on ComfyUI core alone; each other mode pulls in its own optional pack (Audio Drive needs comfyui-vrgamedevgirl, Keyframes/Extend/Chain need ComfyUI-H3-Motion-Context-MultiRef, Mask needs MaskVidExperiments plus the SAM 3.1 checkpoint, and so on - the README's requirements table is exhaustive).

The model files all come from Comfy-Org/MiniMax-H3 into your standard folders:

| File | Folder | |------|--------| | 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.safetensors | vae/ | | minimax_h3_audio_vae_fp32.safetensors | vae/ |

And a word on hardware: this is a 33B omni-modal model with a 32B text encoder. The diffusion models are pruned int8 and the text encoder is 4-bit AWQ, but that's still a serious VRAM bill - this is not a 3060 toy. No verified consumer hardware floor existed at launch, so don't trust the "it'll run on anything" comments.

Troubleshooting

• "Node not found" on a mode you just enabled - that's a missing per-mode pack from the requirements table, not a bug. Install it via ComfyUI Manager, do a full restart, then hard-refresh the browser (Ctrl+F5).
• R2V throws a shape mismatch about broadcasting - a core ComfyUI bug on builds between 2026-08-06 and 2026-08-13, fixed upstream on the 13th. Update ComfyUI.
• Chain or Extend fail on ComfyUI 0.33 - if you have ComfyUI-MiniMaxH3-Cache installed, it patches MiniMaxH3Model._forward in a way that conflicts with Motion Context. Disable that pack (rename the folder to end in .disabled) when you want Chain/Extend.
• Mask mode mixing manual inpainting with a text prompt doesn't work well - the author tested it; pick one approach per mask.

Two things to check before you fall in love: this is beta, and it's developed against a pinned stack (ComfyUI 0.33, listed in COMPATIBILITY.md) - update a dependency and a render may break, so that file is your first stop. And H3's own community license geofences the local weights out of the US, EU, UK and Korea - if you're in one of those regions, the hosted Hailuo API is your only licensed path. For everyone else, this is the least-friction way to run H3 locally.