mmdit

MMDiT

Implémentation d'une seule couche du MMDiT, proposée par Esser et al. dans Stable Diffusion 3, dans Pytorch

Outre une reproduction directe, cela se généralisera également à> 2 modalités, car je peux imaginer un MMDiT pour les images, l'audio et le texte.

Proposera également une variante improvisée d'auto-attention qui sélectionne de manière adaptative les poids à utiliser grâce à un déclenchement appris. Cette idée est venue des convolutions adaptatives appliquées par Kang et al. pour GigaGAN.

Installer

$ pip install mmdit

Usage

 import torch
from mmdit import MMDiTBlock

# define mm dit block

block = MMDiTBlock (
    dim_joint_attn = 512 ,
    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
)

Une version généralisée peut être utilisée ainsi

 import torch
from mmdit . mmdit_generalized_pytorch import MMDiT

mmdit = MMDiT (
    depth = 2 , 
    dim_modalities = ( 768 , 512 , 384 ),
    dim_joint_attn = 512 ,
    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

 @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 }
}

 @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 }
}

 @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 }
}