UPDATE!!! made some changes to the nodes. do a git pull and re-run requirements.
MUST USE NEW TEST WORKFLOW!!!! there are new nodes to unload weight of first model.
if it runs for you, please post your generations in the discussions!!!!! i cant run it, its too big at 30gb of total size with everything.
model details below with files link.
V
EXPERIMENTAL! updates to the nodes are on pause until i can successfully generate with the model. ive made changes to the nodes but i will not commit them to github until the model runs. im sorry if you experienced any issues or wasted time testing the model. i will be sure to update the deacription as soon as it works.
just merged the files to match another contributors workflow. this will lessen the requirements for lower vram users. its still a 16gb fp8 model and the SR model to upscale it heavy as well. working on offloading implementation without sage or triton. it will be in the workflow once i test it. i will update the repo with a SEPERATE offloading workflow check back soon.
THIS MODEL WILL NOT RUN WITHOUT MY CUSTOM NODE SET. ITS CURATED FOR THIS MODEL SPECIFICALLY!
t5gemma text encoder gguf goes in "gguf" folder NOT text encoder folder
files:
https://huggingface.co/realrebelai/DaVinci_MagiHuman_fp8_merges/tree/main
fp8 model - Diffusion models folder
text encoder gguf - gguf folder, NOT text encoder folder
wan vae - vae folder
sd audio vae - vae folder
turbo vae - vae folder
sr fp8 model - diffusion models folder
.png){kind=link}
DaVinci-MagiHuman:Speed by Simplicity: A Single-Stream Architecture for Fast Audio-Video Generative Foundation Model
[fork is based on smthemex's node set: https://github.com/smthemex/ComfyUI_MagiHuman]
-
REBEL added fp8 scaling support for my model, i also added a bypass of flash attn and tiled vae encoding. not tested because i OOM on my 3070 (8gb vram). just doing my part to the community. https://huggingface.co/realrebelai/DaVinci_MagiHuman_fp8_merges/tree/main
-
add layer offload num to fit high vram 大显存 把offload开到你跑得动为止，小显存则从1开始测试，MagiCompiler库不用，但是改了麻烦，还是加了回来
In the ./ComfyUI/custom_nodes directory, run the following:
git clone https://github.com/RealRebelAI/ComfyUI_MagiHuman
pip install -r requirements.txt
#python312以下，注释掉MagiCompiler目录下的pyproject.toml的13行的requires-python = ">=3.12" 再安装
git clone https://github.com/SandAI-org/MagiCompiler.git
cd MagiCompiler
pip install -r requirements.txt
pip install .
- dit and TE links
├── ComfyUI/models/
| ├── diffusion_models/
| ├──fp8.safetensors
| ├──1080p_sr_merge_fp8.safetensors
| ├── vae/
| ├──sd_audio.safetensors
| ├──Wan2.2_VAE.pth
| ├── gguf
| ├──t5gemma-9b-9b-ul2-Q6_K.gguf
We thank the open-source community, and in particular Wan2.2 and Turbo-VAED, for their valuable contributions.
This project is released under the Apache License 2.0.
@misc{davinci-magihuman-2026,
title = {Speed by Simplicity: A Single-Stream Architecture for Fast Audio-Video Generative Foundation Model},
author = {SII-GAIR and Sand.ai},
year = {2026},
url = {https://github.com/GAIR-NLP/daVinci-MagiHuman}
}