Why this matters MiniMax H3 brought open weights for an omni-modal audio+video generator that can produce up to ~15s of 24 FPS video with native stereo audio. This dataset lowers the practical barrier to running H3 locally in ComfyUI by bundling repackaged checkpoints, compatible text encoders and VAEs, and ready-made ComfyUI templates — so you can skip much of the manual repackaging and get to local inference faster.
What Sets It Apart
- ComfyUI-focused packaging: files are repackaged and organized to drop into ComfyUI model folders and templates, plus six official workflow templates (T2V/I2V/R2V and API variants) to run text→video, image→video, and reference→video modes.
- Multiple precision & quant options: offers BF16, INT8, and pruned INT8 variants of the Base checkpoints and matching text-encoder quant formats (NVFP4/AWQ/INT8) so you can trade off VRAM vs fidelity depending on your GPU.
- Joint audio+video artifacts included: video VAE and audio VAE matching the H3 base model are provided so outputs include native stereo dialogue, SFX, and music without a separate audio pipeline.
- Practical file-size notes: checkpoints and encoders are large (tens of GB each); pruned/quantized builds reduce footprint for constrained GPUs.
Who it's for, and tradeoffs
Great fit if you want to run MiniMax H3 locally via ComfyUI and prefer an opinionated, ready-to-drop dataset of model files and templates. It saves time versus manual conversion and offers quantized options for smaller GPUs. Look elsewhere if you need the full H3 pipeline (the Context-IR preprocessing and the 2K Regenerate module are not bundled here), if you lack a capable GPU, or if you require cloud-hosted inference with managed scaling — this distribution targets local inference and experimentation and assumes you can handle large model files and GPU setup.
