| --- |
| license: other |
| license_name: minimax-h3-community-license |
| license_link: https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/LICENSE |
| base_model: MiniMaxAI/MiniMax-H3 |
| tags: |
| - minimax-h3 |
| - fl2va |
| - first-last-frame |
| - video-generation |
| - comfyui |
| - fp8 |
| - float8-e4m3fn |
| --- |
| |
| # MiniMax H3 FL2VA FP8 E4M3FN for ComfyUI |
|
|
| Unofficial mixed-precision FP8 conversion of the MiniMax H3 FL2VA diffusion model for ComfyUI. FL2VA is the MiniMax H3 task checkpoint for text-to-audio-video and native first-frame, last-frame, or first-and-last-frame conditioning. |
|
|
| ## File |
|
|
| - `minimax_h3_fl2va_fp8_e4m3fn.safetensors` |
| - Size: 33,236,581,696 bytes (30.954 GiB) |
| - SHA-256: `E5C2E4390756BB673E44310BECF7317EB262C1BD969936FB50AC64E7774A73AF` |
|
|
| ## Conversion |
|
|
| The official BF16 checkpoint was converted tensor-by-tensor without materializing the complete checkpoint in RAM or VRAM. Multi-dimensional BF16 weight tensors were converted to FP8 E4M3FN. Normalization, bias, conditioning, final-layer, and time-embedding tensors retained BF16 or FP32 precision. |
|
|
| - Total tensors: 535 |
| - FP8 E4M3FN tensors: 258 |
| - BF16 tensors: 264 |
| - FP32 tensors: 13 |
| - Source file: `minimax_h3_fl2va_bf16.safetensors` |
| - Source size: 66,280,487,368 bytes |
| - Source SHA-256: `907D4ADD438438EC1544F5240C3B38532ED934FE6BE75677A6BBDA2A6FDD6182` |
|
|
| Tensor names and shapes were checked against the source checkpoint and match exactly. The output safetensors data ranges cover the complete file without gaps or truncation. |
|
|
| ## ComfyUI |
|
|
| Place the checkpoint in `ComfyUI/models/diffusion_models/` and load it with the standard `UNETLoader` using `weight_dtype: default`. |
|
|
| For native first/last-frame control, use the core `MiniMaxH3ImageToVideo` node and connect images to its `first_frame` and/or `last_frame` inputs. This FL2VA checkpoint is different from the Ref2VA checkpoint used by `MiniMaxH3ReferenceToVideo`; it does not add Ref2VA multi-reference conditioning. |
|
|
| Verified locally with ComfyUI 0.29.0 and an NVIDIA RTX 5090 using the core `MiniMaxH3ImageToVideo` node with both `first_frame` and `last_frame` connected: |
|
|
| - 768 x 1344 |
| - 124 frames at 24 fps |
| - 2-step full-resolution smoke test |
| - `res_multistep` sampler |
| - `beta` scheduler |
| - `UNETLoader` with `weight_dtype: default` |
|
|
| ComfyUI completed model loading, first/last keyframe VAE conditioning, sampling, video/audio VAE decoding, and MP4 encoding successfully. The resulting file contained 124 H.264 frames at 24 fps and 32 kHz stereo AAC audio. The two-step setting verifies execution and is not a production-quality recommendation; use the official 20-step setting for final output. |
|
|
| ## License and source |
|
|
| Base model: [MiniMaxAI/MiniMax-H3](https://huggingface.co/MiniMaxAI/MiniMax-H3) |
|
|
| This is a modified derivative and is not an official MiniMax release. This repository includes the upstream MiniMax H3 Community License Agreement and required `NOTICE`. The upstream license contains territorial restrictions and excludes the European Union, United Kingdom, Republic of Korea, and United States from the applicable territory unless separately authorized by MiniMax. Review the full `LICENSE` before downloading, using, or redistributing this checkpoint. |
|
|
