https://github.com/lucidrains/mmdit
Implementation of a single layer of the MMDiT, proposed in Stable Diffusion 3, in Pytorch
https://github.com/lucidrains/mmdit
artificial-intelligence attention-mechanisms deep-learning multi-modal-attention
Last synced: 5 months ago
JSON representation
Implementation of a single layer of the MMDiT, proposed in Stable Diffusion 3, in Pytorch
- Host: GitHub
- URL: https://github.com/lucidrains/mmdit
- Owner: lucidrains
- License: mit
- Created: 2024-05-04T14:38:14.000Z (about 2 years ago)
- Default Branch: main
- Last Pushed: 2025-01-12T18:04:29.000Z (over 1 year ago)
- Last Synced: 2025-09-25T02:22:11.863Z (9 months ago)
- Topics: artificial-intelligence, attention-mechanisms, deep-learning, multi-modal-attention
- Language: Python
- Homepage:
- Size: 156 KB
- Stars: 449
- Watchers: 3
- Forks: 12
- Open Issues: 3
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
## MMDiT
Implementation of a single layer of the MMDiT, proposed by Esser et al. in Stable Diffusion 3, in Pytorch
Besides a straight reproduction, will also generalize to > 2 modalities, as I can envision an MMDiT for images, audio, and text.
Will also offer an improvised variant of self attention that adaptively selects the weights to use through learned gating. This idea came from adaptive convolutions applied by Kang et al. for GigaGAN.
## Install
```bash
$ pip install mmdit
```
## Usage
```python
import torch
from mmdit import MMDiTBlock
# define mm dit block
block = MMDiTBlock(
dim_cond = 256,
dim_text = 768,
dim_image = 512,
qk_rmsnorm = True
)
# mock inputs
time_cond = torch.randn(2, 256)
text_tokens = torch.randn(2, 512, 768)
text_mask = torch.ones((2, 512)).bool()
image_tokens = torch.randn(2, 1024, 512)
# single block forward
text_tokens_next, image_tokens_next = block(
time_cond = time_cond,
text_tokens = text_tokens,
text_mask = text_mask,
image_tokens = image_tokens
)
```
A generalized version can be used as so
```python
import torch
from mmdit.mmdit_generalized_pytorch import MMDiT
mmdit = MMDiT(
depth = 2,
dim_modalities = (768, 512, 384),
dim_cond = 256,
qk_rmsnorm = True
)
# mock inputs
time_cond = torch.randn(2, 256)
text_tokens = torch.randn(2, 512, 768)
text_mask = torch.ones((2, 512)).bool()
video_tokens = torch.randn(2, 1024, 512)
audio_tokens = torch.randn(2, 256, 384)
# forward
text_tokens, video_tokens, audio_tokens = mmdit(
modality_tokens = (text_tokens, video_tokens, audio_tokens),
modality_masks = (text_mask, None, None),
time_cond = time_cond,
)
```
## Citations
```bibtex
@article{Esser2024ScalingRF,
title = {Scaling Rectified Flow Transformers for High-Resolution Image Synthesis},
author = {Patrick Esser and Sumith Kulal and A. Blattmann and Rahim Entezari and Jonas Muller and Harry Saini and Yam Levi and Dominik Lorenz and Axel Sauer and Frederic Boesel and Dustin Podell and Tim Dockhorn and Zion English and Kyle Lacey and Alex Goodwin and Yannik Marek and Robin Rombach},
journal = {ArXiv},
year = {2024},
volume = {abs/2403.03206},
url = {https://api.semanticscholar.org/CorpusID:268247980}
}
```
```bibtex
@inproceedings{Darcet2023VisionTN,
title = {Vision Transformers Need Registers},
author = {Timoth'ee Darcet and Maxime Oquab and Julien Mairal and Piotr Bojanowski},
year = {2023},
url = {https://api.semanticscholar.org/CorpusID:263134283}
}
```
```bibtex
@article{Zhu2024HyperConnections,
title = {Hyper-Connections},
author = {Defa Zhu and Hongzhi Huang and Zihao Huang and Yutao Zeng and Yunyao Mao and Banggu Wu and Qiyang Min and Xun Zhou},
journal = {ArXiv},
year = {2024},
volume = {abs/2409.19606},
url = {https://api.semanticscholar.org/CorpusID:272987528}
}
```