File size: 4,793 Bytes
21665fa | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 | # 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.
|