ComfyUI ALL-in-ONE MiniMax H3: one node for the whole H3 video pipeline

The graph surgery this node is meant to remove

MiniMax H3 work in ComfyUI normally arrives as a set of official native workflows, plus community packs for the parts those workflows do not cover. Getting from a text prompt to a finished clip with audio means locating the right workflow, loading it, and understanding which loaders and samplers feed which stage. Doing reference-driven video, or continuing an existing clip, means more of that. The README frames the problem bluntly: "No node graph to build, no wires to connect, no hunting through twelve custom node packs to figure out which workflow is the right one."

The intended user is someone who already has H3 models and the surrounding ComfyUI environment in place and wants to spend time on prompts and references rather than on graph topology. It is not a way to run H3 without the models. T2V, I2V and R2V are stated to run on ComfyUI core alone, but everything else depends on model files and optional custom node packs listed in requirements/REQUIREMENTS.md. The node is a front end over that stack, not a replacement for it.

One node, eleven modes, and what each mode actually calls

The mechanism is a mode switch inside a single node. The README's mode table maps each entry to a distinct pipeline: Image covers text to image, image edit, and reference mix with up to nine references through H3 Studio; T2V produces video with native audio using the fl2va model; I2V animates a start frame and can morph to an end frame; R2V lets reference images, videos or audio drive the clip through the ref2va model. Character Sheet stitches up to nine reference images of one character into a six-panel full orbit or a faster four-panel turnaround. Audio Drive uses your track as the soundtrack and drives mouth movement. Keyframes pins still images at arbitrary frame positions. Extend continues an existing video. Chain does multi-clip continuation through H3 Motion Context on the latent path with no re-encode. Mask paints or describes a region, tracks it, and replaces only that area. Upscale is described as an RTX/Seed2VR Video Super Resolution hook.

The credit list shows how much of this is assembled rather than written from scratch. Chain, Keyframes and Extend wiring comes from ComfyUI-H3-Motion-Context-MultiRef by seitanism, which is also why pyproject.toml ships GPL-3.0. Mask mode uses crop and latent-mask nodes from MaskVidExperiments. Image mode uses ComfyUI-MiniMax-H3-Studio. The base graphs are the official MiniMax H3 native workflows from Comfy-Org. The "one node" idea itself is credited to Ján's one-node-flux-2-klein and one-node-gemma-4 projects. That is a lot of upstream surface behind one button, and it is the main reason the compatibility document exists.

Installing the node and running a first text-to-video job

Installation is a clone into the custom_nodes directory. The README gives exactly one command, with no pip step and no extra flags:

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

After the clone, restart ComfyUI. The README then says to double-click the canvas and search for ALL in ONE MiniMaxH3. That search is how you add the node; there is no separate menu entry described.

The first job worth running is the one that needs the least: T2V on core. Switch the node to T2V, type a prompt, leave the optional presets alone, and generate. T2V is stated to run on ComfyUI core alone, so a failure here points at your ComfyUI install or your model files rather than at a missing custom node pack. If you want the fast prompt-tweak preset, the README notes that SLA Draft needs the ComfyUI-PlagueKind-Nodes pack and is off by default, so leave it off for a first run.

Once T2V works, the other modes are gated by the requirements document rather than by the node itself. Read requirements/REQUIREMENTS.md before switching to Mask, Character Sheet, or Audio Drive, and expect to install model files and additional node packs for those paths.

Compare and Stitch, and the 24 fps frame matching

One feature sits outside the mode list. Compare & Stitch takes two to four outputs from the Library or the outputs strip. The Compare tab plays every clip side by side in sync. The Stitch tab exports them as one side-by-side clip with deterministic frame matching at 24 fps, configurable spacing and gutter color, per-clip trim, and h264 or h265/mp4 output. Audio comes from clip 1. The README states this needs only ComfyUI VideoHelperSuite, which the node already lists.

Deterministic frame matching at a fixed 24 fps is the detail worth noticing. Clips generated at other frame rates or with dropped frames will be conformed to that grid rather than preserved at their native timing, and the audio rule means clips 2 through 4 contribute picture only. If your comparison depends on each clip's own soundtrack, this is the wrong tool for that job. The README does not describe how trim values are expressed, so check the node's own controls before assuming a unit.

Mask mode: pick painting or prompting, not both

Mask mode is the clearest example of a documented limitation. The README states that painting a region and describing it with text are two separate paths, and that in the author's testing, mixing manual inpainting with a text prompt does not work well. The instruction is to pick one approach per mask.

That is a real constraint on how you work, not a note about polish. It means the two natural workflows, rough paint plus a text description of what belongs there, and a purely described region, do not combine. Plan a shot around one of them. The masking demo asset is named Sam3Preview.gif, which suggests a SAM-family tracking component, but the README does not name the tracker or its model, so treat the tracking behaviour as something to verify on your own footage rather than something the documentation pins down.

Compatibility is pinned, and the project ships no releases

The README says the author develops and tests against a pinned stack: a ComfyUI version, custom node commits, and model files, all listed in COMPATIBILITY.md. The advice for a failed render after an update is to start there. This is the honest shape of a project that sits on top of several moving upstreams, and it means an upgrade of ComfyUI or of any credited pack can break a mode without anything changing in this repository.

The repository has no releases retrieved, and pyproject.toml carries version 0.20.0. So there are no tagged versions to pin against; the practical pin is a commit hash of the repository itself, alongside the versions in COMPATIBILITY.md. The README also carries a beta status badge and asks users to open an issue when something breaks, and it does not document a rollback path. If you need a dependency you can freeze and reason about, that absence matters more than any single mode.

On licensing: the project is GPL-3.0, and pyproject.toml explains why, stating that the Motion Context, keyframes and extension wiring is derived from the GPL-3.0 ComfyUI-H3-Motion-Context-MultiRef fork, so the pack ships GPL-3.0. If you distribute a workflow or a bundle that includes this code, that is the licence you are working under. Whether your specific use counts as distribution is a question for a lawyer, not for this article.

Where a hand-built graph still wins

The alternative is building the graph yourself from the official MiniMax H3 native workflows from Comfy-Org, adding ComfyUI-H3-Motion-Context-MultiRef for chains, keyframes and extension, ComfyUI-MiniMax-H3-Studio for image work, and MaskVidExperiments for masking. That is the same set of components this node wraps.

The difference is who owns the wiring. In a hand-built graph you can replace one stage, swap a sampler, or keep two versions of a pipeline side by side in separate workflows. You can also see exactly which node produced a bad frame. With the all-in-one node you get one panel and a mode switch, at the cost of that visibility: when a render fails, the first question is which upstream pack or model is at fault, and the answer lives in COMPATIBILITY.md rather than in your own graph. For a single person iterating on prompts, the trade is usually worth it. For a team maintaining several pipelines that share components, it usually is not.

Editorial conclusion

Adopt it if you already run MiniMax H3 workflows in ComfyUI and want the graph wiring handled for you, especially for Chain, Keyframes and Extend, which the README credits to a separate Motion Context fork. Skip it if you need a stable, versioned dependency: the project ships no releases, its own badge says beta, and the README does not document rollback. Before you rely on it, open COMPATIBILITY.md and requirements/REQUIREMENTS.md and check that your ComfyUI version, custom node commits and model files match the pinned stack the author develops against.

Frequently asked questions

Official sources

1. Issues
2. LeonQ8/ComfyUI-ALLinONE-MinimaxH3 on GitHub
3. License: GPL-3.0
4. README