| # Modified files |
|
|
| Section III.2 of the MiniMax H3 Community License Agreement requires that modified |
| files carry a prominent notice saying so. This file is that notice. |
|
|
| Everything below is derived from |
| [`MiniMaxAI/MiniMax-H3`](https://huggingface.co/MiniMaxAI/MiniMax-H3). |
|
|
| ## Modified: the base `transformer/`, in `int8/` and `gguf/` |
|
|
| The eight files in `int8/` and `gguf/` are **modified copies of the base model's `transformer/`** |
| (the FL2VA transformer). Each one is that transformer with one of the two LoRAs below merged into its |
| weights at fp32, then quantized: |
|
|
| - `int8/viggle_remake_{swap,edit}_pruned_int8_convrot.safetensors`: the merged transformer in |
| ComfyUI's int8 format, following the layout of Comfy-Org's repackaged |
| `minimax_h3_fl2va_pruned_int8_convrot` (itself derived from MiniMax H3). The attention and MLP |
| linears are int8 with per-row scales; adaLN is that file's `pruned` form, unchanged. |
| - `gguf/viggle_remake_{swap,edit}-{Q8_0,Q6_K,Q4_K_M}.gguf`: the merged transformer with the |
| attention and MLP linears quantized to GGUF types, and the `wan` architecture tag so that |
| ComfyUI-GGUF loads it. |
|
|
| Changes against the base `transformer/`: the LoRA deltas added to 302 linear layers, the |
| quantization, tensor names repacked to ComfyUI's layout, and the adaLN pruning above. Nothing else. |
|
|
| ## Not modified |
|
|
| The two LoRAs below and the ComfyUI LoRA files are deltas that apply to the base model's |
| `transformer/` at load time. That `transformer/` stays in your own copy of the base model and is |
| not changed on disk. |
|
|
| ## `lora/swap/pytorch_lora_weights.safetensors` — new |
|
|
| Not a MiniMax file. Made by us. A PEFT LoRA, rank 256 and alpha 256, over 302 linear |
| layers of the base model's `transformer/` (the FL2VA transformer), for replacing people |
| and characters. It is stored in bf16, except for the 4 tensors on `proj_in` and |
| `proj_out`, which are fp32. |
|
|
| It is a delta on the stock base `transformer/`. Load it with |
| `load_lora_adapter(..., prefix=None)` and put the result in the `transformer_ref` slot of |
| the `ref2va` workflow, with keyframes pinned inside the generated span. We have not run it on any other base, including the base |
| model's own `transformer_ref/`. |
|
|
| ## `lora/edit/pytorch_lora_weights.safetensors` — new |
|
|
| Not a MiniMax file. Made by us. Same shape, base and loading as `lora/swap/`, for |
| general video edits: object swap, removal, insertion, background, weather and light. |
|
|
| ## `lora_lowrank/{swap,edit}-r{128,64}/pytorch_lora_weights.safetensors` — new |
| |
| Not MiniMax files. Made by us. The two LoRAs above cut to rank 128 or 64 (alpha = rank): each |
| layer's delta `B @ A` is replaced by its best rank-k approximation from an SVD, and nothing else |
| changes. Same base, layers and loading as `lora/swap/` and `lora/edit/`; `inference/sample.py |
| --rank 128` or `--rank 64` picks them. |
| |
| ## `comfyui/` — new |
| |
| Not MiniMax files. Made by us. `viggle_remake_{swap,edit}_comfyui.safetensors` hold the same LoRA |
| weights as `lora/swap` and `lora/edit`, repacked to ComfyUI's layer names by `comfyui/to_comfyui.py`. |
| The custom nodes and workflows are code. |
|
|
| ## `inference/`, `examples/demo.sh`, `examples/prompts/` — new |
|
|
| Not MiniMax files. Written by us. The prompts follow the six-section format that |
| MiniMax documents in the base model's `docs/`. |
|
|
| ## `LICENSE-CODE`, `NOTICE` — new |
|
|
| Not MiniMax files. `LICENSE-CODE` is the Apache 2.0 text, and it covers the code in `inference/`, `examples/` and |
| `comfyui/` (the `.py` files and the workflow `.json` files) only. Each script there carries an `SPDX-License-Identifier: Apache-2.0` |
| header. |
|
|
| `NOTICE` records that the weights are *not* Apache 2.0. |
|
|
| `LICENSE` is the MiniMax H3 Community License Agreement itself, included unmodified as |
| Section III.1 requires. |
|
|
| ## `examples/media/`, `examples/inputs/` — new |
|
|
| Not MiniMax files. They show third-party footage rendered by this model. Most clips in |
| `examples/media/` are side-by-sides, the source on the left and the render on the right; |
| `white-house-funk.mp4` is the render alone. `examples/inputs/` holds two source spans at |
| 832 × 480 with their soundtracks, plus the frames we painted over them with OpenAI's |
| `gpt-image` models. |
|
|
| The footage is credited in the "Example footage" section of `README.md`. Neither |
| `LICENSE` nor `LICENSE-CODE` grants any rights to it. |
|
|
| ## Not redistributed here |
|
|
| The following are **not** included in this repository and are not modified: |
|
|
| - the unmodified transformers, `transformer/` and `transformer_ref/` (the merged, quantized copies in |
| `int8/` and `gguf/` are listed above); |
| - the text encoder; |
| - the VAE and audio VAE; |
| - the schedulers; |
| - the tokenizer and processor. |
|
|
| They are loaded at runtime from your own copy of `MiniMaxAI/MiniMax-H3`, which you must |
| download separately and under its own licence terms. |
|
|
