ComfyUI_MiniMaxH3_Director: five MiniMax H3 workflows you can import
ComfyUI MiniMax H3 Director workflow
At a glance
- What is it?
- This repository is a mirror of the upstream MiniMax H3 Director node, holding five importable ComfyUI workflow JSON files for text-to-video, first and last frame, reference-to-video, video editing and video reference editing. It is a download and reproduction copy, not the plugin itself.
- Who is it for?
- Adopt this repository if you already run ComfyUI 0.30.0 or newer and have the MiniMax H3 INT8 weights, and you want five ready-made graphs rather than building the graph yourself. Skip it if you are on macOS, since the verified environment is an RTX 4090 48GB, or if you expect a plugin: the node code lives upstream at AIMixer/ComfyUI_MiniMaxH3_Director.
- Can I use it commercially?
- Yes. Apache-2.0 is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 45 days ago.
- What is it written in?
- GitHub does not report a main language for this repository.
Answers come from the project's GitHub data, last synced on September 16, 2026, and from our analysis. They are not legal advice.
DEEP OPEN-SOURCE ANALYSIS
What the MiniMax H3 Director workflow repository actually contains
The repository is small by design. Its top-level entries are LICENSE, README.md and example_workflows/. Inside example_workflows/ sit five JSON files, each a complete ComfyUI graph: minimax_h3_director_t2v.json, minimax_h3_director_fl2v.json, minimax_h3_director_r2v.json, minimax_h3_director_v2v.json and minimax_h3_director_rv2v.json. The README describes this as a copy kept for downloading, reproducing and testing, and states plainly that the workflows come from AIMixer/ComfyUI_MiniMaxH3_Director.
That framing matters when you decide whether to adopt it. This is not the node implementation. It is the graph layer: the wiring, node settings and model loader choices that make MiniMax H3 usable in ComfyUI without assembling everything by hand. The audience is the person who has already decided to run MiniMax H3 locally and now needs a starting point that someone else has actually opened in ComfyUI. The README lists a verified environment of an NVIDIA RTX 4090 48GB, ComfyUI 0.30.0, PyTorch 2.11.0 with CUDA 12.8, and MiniMax H3 Ref2VA INT8, with the MiniMaxH3Director node registering successfully.
The five modes and which checkpoint each one loads
The workflows split by mode, and the split is not cosmetic: each mode expects a different UNET file. T2V, FL2V and I2V all use the fl2va checkpoint. The README gives its path as models/diffusion_models/minimax_h3_fl2va_pruned_int8_convrot.safetensors. R2V, V2V and RV2V use ref2va instead, at models/diffusion_models/minimax_h3_ref2va_pruned_int8_convrot.safetensors.
What each mode is for, per the README table: T2V generates audio and video from text. FL2V takes a first and last frame, and doubles as I2V when you supply only the first frame. R2V generates from image, video and audio references. V2V edits a source video segment by segment along its timeline. RV2V combines a source video with a reference image or reference audio, and the README names video face replacement as its fitting test case.
Three components are shared across all five: models/text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors, models/vae/minimax_h3_video_vae_fp16.safetensors and models/vae/minimax_h3_audio_vae_fp32.safetensors. The README also specifies that the CLIP Loader type must be set to minimax. This is the kind of detail that silently breaks a graph when missed, and it is the main reason a working JSON is worth more than a description of one.
Installing the upstream node and importing your first workflow
The repository itself is not installed. You clone the upstream plugin into ComfyUI's custom_nodes directory, install its requirements, restart ComfyUI, then drag a JSON from example_workflows/ onto the page. The README requires ComfyUI 0.30.0 or higher.
cd ComfyUI/custom_nodes
git clone https://github.com/AIMixer/ComfyUI_MiniMaxH3_Director.git
python -m pip install -r ComfyUI_MiniMaxH3_Director/requirements.txtAfter the restart, the MiniMaxH3Director node should appear in the node list. If it does not, the README's verified environment note is the reference point: the node registered successfully under ComfyUI 0.30.0 with PyTorch 2.11.0 and CUDA 12.8.
For a first real run, pick the mode your weights support. The README states that the verified 4090 environment only had Ref2VA, so R2V, V2V and RV2V are the ones that can run directly there. T2V, I2V and FL2V need the fl2va weights added first.
models/diffusion_models/minimax_h3_ref2va_pruned_int8_convrot.safetensors
models/text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors
models/vae/minimax_h3_video_vae_fp16.safetensors
models/vae/minimax_h3_audio_vae_fp32.safetensorsWith those in place, import example_workflows/minimax_h3_director_r2v.json and set the CLIP Loader type to minimax.
Segment-by-segment control and the RV2V face replacement loop
The director node's practical value is that it treats a long clip as segments rather than one submission. The README says the node supports per-segment running and caching, so changing one segment does not force a rerun of the whole timeline. Audio can be model-generated, carried over from the source, or muted, per segment.
RV2V is the most involved flow and the README lays it out as five steps. Import example_workflows/minimax_h3_director_rv2v.json. Upload the source video and split it by shot, either manually or with smart segmentation. Upload the person reference image, adding reference audio if you need voice constraints. Write a prompt per segment, where the source clip binds as <Video 1> and references use <Picture N> and <Audio J>. Then generate five seconds first and inspect face, clothing, motion and shot boundaries before extending.
The five-second check is the honest part of this document. It tells you the boundary between a segment is handed off from the previous segment's last frame to the next segment's first frame, and states that this handoff cannot substitute for a person consistency check. In other words, the graph gives you continuity of frames, not continuity of identity.
Where the MiniMax H3 Director workflow falls short
There is no hard identity lock. The README says the person's identity is constrained jointly by the model and the reference material, and that the director node does not impose a hard identity lock on top of that. If your use case requires a face to stay identical across many segments with no drift, this is the wrong tool to rely on unassisted.
The verified environment is narrow. One GPU, one ComfyUI version, one weight variant. The README documents a single configuration, so anything outside an RTX 4090 48GB running ComfyUI 0.30.0 with the INT8 checkpoints is unverified territory. There is no macOS path described anywhere in the repository, and no CPU fallback mentioned.
Model weights are a separate dependency with separate terms. The README notes that the workflows and the upstream plugin ship under Apache-2.0, but that the model weights follow their own licences. Downloading the JSON files does not give you the right to use the checkpoints under Apache-2.0.
Finally, the repository is a mirror. Bug reports and node fixes belong upstream. If the node misbehaves, the copy here will not diverge to fix it.
How this differs from building the graph yourself, and from other ComfyUI video stacks
The real alternative is not another product. It is the upstream repository at AIMixer/ComfyUI_MiniMaxH3_Director, which contains the node code and, per this README, is where the workflows originate. If you clone upstream directly, you get the same node and you can follow its own examples. What you lose is the curated set: five JSON files organised by mode, with a table mapping each file to its UNET and its purpose, plus a verified environment list and an RV2V walkthrough in one place. What you gain is that fixes and issues land where the code lives.
A second alternative is to build the graph in ComfyUI from scratch. That is a legitimate choice if you want to understand every connection, and it avoids depending on a JSON whose node versions may drift from your installed plugin. The cost is that you have to rediscover the model paths, the CLIP Loader type setting and the segment handoff behaviour yourself. The README's own A/B testing note applies either way: hold source material, prompt, seed, resolution, frame count and steps fixed when comparing runs.
Licence, maintenance and the cost of keeping up
The workflows and the upstream plugin are released under Apache-2.0, with the LICENSE file at the repository root. That is a permissive licence, but the README explicitly separates it from the weights: the model weights follow their own terms. If you plan to redistribute anything you generate or any modified graph, read the weight licences rather than assuming Apache-2.0 covers the whole stack. This is a description of what the README states, not legal advice.
The last push to this repository was on 2026-08-04. There are no releases listed. The upgrade surface is therefore the upstream plugin and the ComfyUI version it targets, not this repository. The README pins its verification to ComfyUI 0.30.0, and it requires 0.30.0 or higher for installation, so a ComfyUI upgrade is the event most likely to invalidate the graph. SageAttention is optional: the README says to place the patch node between the UNETLoader and the model input of MiniMaxH3Director, and to confirm output quality is unchanged before recording any speed difference.
Editorial conclusion
Adopt this repository if you already run ComfyUI 0.30.0 or newer and have the MiniMax H3 INT8 weights, and you want five ready-made graphs rather than building the graph yourself. Skip it if you are on macOS, since the verified environment is an RTX 4090 48GB, or if you expect a plugin: the node code lives upstream at AIMixer/ComfyUI_MiniMaxH3_Director. Before committing, confirm you can register the MiniMaxH3Director node and that the fl2va or ref2va checkpoint matching your chosen workflow is present in models/diffusion_models.
Frequently asked questions
What is MiniMax AI used for?
The README does not describe MiniMax AI as a general service. It covers MiniMax H3 running locally in ComfyUI, where the five workflows generate video and audio from text, first and last frames, reference images, reference video and reference audio.
Is MiniMax H3 on ComfyUI?
Yes. The README requires ComfyUI 0.30.0 or higher and the upstream AIMixer/ComfyUI_MiniMaxH3_Director node, which registers as MiniMaxH3Director. The repository supplies five importable workflow JSON files for it.
How to use MiniMax H3 locally?
Clone AIMixer/ComfyUI_MiniMaxH3_Director into ComfyUI/custom_nodes, install its requirements.txt, restart ComfyUI, then drag a JSON from example_workflows/ onto the page. You also need the matching diffusion model, the qwen3vl text encoder and both VAE files in the paths the README lists, with the CLIP Loader type set to minimax.
How to use MiniMax AI for free?
The README does not describe a free tier or hosted service. It covers local execution in ComfyUI with downloaded INT8 checkpoints, and notes that the weights follow their own licence terms separate from the Apache-2.0 licence on the workflows and plugin.
Community notes