The community minimax-h3-comfyui-video plugin lets an LM Studio chat model request a short MiniMax H3 video through ComfyUI. The language model writes the prompt and calls a tool; ComfyUI performs video generation, and the plugin returns a preview and an MP4 link.
This is a local integration between two running applications. Installing the plugin does not turn a language model into a video model or remove the need for H3 checkpoints. The following setup guide is based on the repository checked October 18, 2026. We have reviewed its documented contract, but have not installed it or repeated its Apple Silicon demos.
Select the right video tool
The repository exposes three tools with different conditioning. Pick the tool according to what the image means, rather than treating all image inputs as interchangeable.
| Tool | Input meaning | Documented model route |
|---|---|---|
minimax_generate_video | A text description creates a clip | FL2VA GGUF with eight-step Turbo LoRA |
minimax_animate_image | The supplied image fixes the first frame; an optional second image fixes the last | FL2VA GGUF with eight-step Turbo LoRA |
minimax_reference_video | One to three images guide a new scene's subject or style | Ref2VA GGUF with four-step Turbo LoRA |
For example, use animation when a robot must begin beside a particular bonsai in exactly the supplied composition. Use reference generation when that robot should appear in a different street scene. A reference image does not promise that its original composition becomes frame one.
The current interface accepts three-to-five-second requests. Its landscape, portrait, and square canvases are approximately 0.4 megapixels. Duration follows H3's frame grid: the documented five-second request becomes 124 frames at 24fps, approximately 5.17 seconds. Check the exported file if an editor requires an exact five-second slot.
Make ComfyUI work before adding chat
The plugin expects ComfyUI at http://127.0.0.1:8188 unless COMFYUI_URL overrides it. Establish a successful small render in that server before diagnosing tool calls. Otherwise a missing checkpoint, loader error, and language-model tool-selection error all appear at once.
The documented custom-node prerequisites are ComfyUI-GGUF and ComfyUI-H3-Multishot. The latter supplies GGUF-aware H3 model and encoder loaders and the architecture patch required by this configuration. ComfyUI 0.30.0 or newer supplies native H3 conditioning and audio/video assembly nodes.
The repository expects separate FL2VA and Ref2VA GGUF diffusion models, a GGUF encoder plus its F16 projector, an FP16 video VAE, an FP32 audio VAE, and the corresponding eight-step or four-step Turbo adapter. Put diffusion models in models/diffusion_models, encoder files in models/text_encoders, VAEs in models/vae, and adapters in models/loras.
Use the exact filenames in the repository's prerequisite table and saved workflows. A similar filename is not sufficient if its architecture, precision, loader, or conditioning path differs. Download only the files needed for the tool you intend to evaluate first.
Build the plugin and choose a preset
Clone the community repository and follow its build instructions from that source checkout. It documents installing dependencies and then running lms dev -i -y from the repository, rather than from an installed plugin folder. After source changes, reload or reinstall through the same development command so LM Studio uses the updated bundle.
Import the companion preset matching your task: Video Director, Video Animator, or Video Reference. These presets teach prompt structure and the response format needed to surface the result in chat. Your chosen language model must also call tools reliably; an attractive natural-language answer does not prove a generation request reached ComfyUI.
For a first text test, ask for one simple subject performing one action, with a short duration and a specified sound. For example: a small wooden boat rocks gently on a pond while water laps against its hull, with no dialogue. This is an editorial test suggestion, not a prompt we have rendered or a guaranteed result.
Inspect where the result goes
The plugin documents an MP4 in LM Studio's working directory and a second copy under ~/Desktop/h3-clips, adjustable with H3_CLIPS_DIR. Its response includes a first-frame thumbnail, a playback link, and the saved path. The thumbnail uses FFmpeg; without it, the clip and path can still be returned.
If a tool appears successful but playback is absent, inspect the saved MP4 before rerunning. This separates a display problem from a failed video job and avoids creating duplicate takes. Confirm that the MP4 contains the expected native audio; a silent GIF in the repository is only a preview format.
Input images must be in allowed directories. The plugin accepts its working directory and documented user folders, with additional paths configurable through H3_ALLOWED_IMAGE_DIRS. A rejected path may therefore be an input-location problem rather than an unsupported image.
Treat Mac timing and launch scripts as local configuration
The author reports M1 Max demos and several-minute-to-tens-of-minutes renders. Those observations are not RTX benchmarks or a universal Mac memory requirement. The README mentions roughly 42GB resident during its render configuration and a 45-minute tool timeout.
Its default auto-launch path points to a local optimized ComfyUI shell script. Prepare your own working launcher and set COMFYUI_LAUNCHER when that path does not exist. A fallback environment setting is not proof that every operation and checkpoint runs natively on MPS.
Start with one text clip, then one first-frame clip, and finally reference generation. Keep a successful workflow and output between stages. For a separate native Mac application, see H3cApp; for the underlying model graph, see the H3 ComfyUI guide. Neither is the same product as this LM Studio plugin.