Transfusion - Pytorch
Pytorch implementation of Transfusion, "Predict the Next Token and Diffuse Images with One Multi-Modal Model", from MetaAI.
In this repo, we will substitute diffusion with flow matching given the success of Flux from Black Forest Labs (but will keep the original paper title given Transflow does not have the same ring). This repository will also attempt to extend to any number of modalities.
Appreciation
- Pranoy for adding classifier free guidance!
Install
$ pip install transfusion-pytorch
Usage
One modality, say images
from torch import randint, randn
from transfusion_pytorch import Transfusion
model = Transfusion(
num_text_tokens = 256,
dim_latent = 384,
modality_default_shape = (4,), # fallback, in the case the language model did not produce a valid modality shape
transformer = dict(
dim = 512,
depth = 8
)
)
# any torch.long is text, torch.float is modalities
text_and_images = [
[randint(0, 256, (16,)), randn(4, 384), randint(0, 256, (8,)), randn(6, 384)],
[randint(0, 256, (16,)), randn(7, 384), randint(0, 256, (5,)), randn(2, 384), randint(0, 256, (9,))]
]
loss = model(text_and_images)
loss.backward()
# after much training
one_multimodal_sample = model.sample()
Multiple different modalities
from torch import randint, randn
from transfusion_pytorch import Transfusion
model = Transfusion(
num_text_tokens = 256,
dim_latent = (384, 192), # specify multiple latent dimensions
modality_default_shape = ((4,), (2,)), # default shapes for first and second modality
transformer = dict(
dim = 512,
depth = 8
)
)
# then for the Tensors of type float, you can pass a tuple[int, Tensor] and specify the modality index in the first position
# any torch.long is text, torch.float is modalities
text_images_and_audio = [
[randint(0, 256, (16,)), (0, randn(4, 384)), randint(0, 256, (8,)), (1, randn(6, 192))],
[randint(0, 256, (16,)), randn(7, 384), randint(0, 256, (5,)), (1, randn(2, 192)), randint(0, 256, (9,))]
]
loss = model(text_images_and_audio)
loss.backward()
# after much training
one_multimodal_sample = model.sample()
Automatically taking care of encoding and decoding of images
import torch
from torch import nn, randint, randn
from transfusion_pytorch import Transfusion, print_modality_sample
mock_encoder = nn.Conv2d(3, 384, 3, padding = 1)
mock_decoder = nn.Conv2d(384, 3, 3, padding = 1)
model = Transfusion(
num_text_tokens = 12,
dim_latent = 384,
channel_first_latent = True,
modality_default_shape = (4, 4),
modality_encoder = mock_encoder,
modality_decoder = mock_decoder,
transformer = dict(
dim = 512,
depth = 8
)
)
text_and_images = [
[
randint(0, 12, (16,)), # 16 text tokens
randn(3, 8, 8), # (8 x 8) 3 channeled image
randint(0, 12, (8,)), # 8 text tokens
randn(3, 7, 7) # (7 x 7) 3 channeled image
],
[
randint(0, 12, (16,)), # 16 text tokens
randn(3, 8, 5), # (8 x 5) 3 channeled image
randint(0, 12, (5,)), # 5 text tokens
randn(3, 2, 16), # (2 x 16) 3 channeled image
randint(0, 12, (9,)) # 9 text tokens
]
]
loss = model(text_and_images)
loss.backward()
# after much training
one_multimodal_sample = model.sample()
print_modality_sample(one_multimodal_sample)
Sampling a batch of interleaved text + modality samples in parallel, with sample_many
sample_many decodes an entire batch of samples concurrently: all samples currently decoding text
share a single kv-cached forward pass, and all samples currently decoding a modality share a single
joint odeint trajectory (one forward pass per ode evaluation, with each sample keeping its own
shape, length and modality type). sample_one walks each sample through the same state machine
serially - sample_many is the batched equivalent.
import torch
from transfusion_pytorch import Transfusion
model = Transfusion(
num_text_tokens = 256,
dim_latent = 384,
modality_default_shape = (4,),
transformer = dict(
dim = 512,
depth = 8
)
)
# a batch is a list of prompts, each taking the same form as the `prompt` of `sample_one` -
# a list of text tensors / (modality_type, tensor) tuples, a raw tensor or tuple, or `None`
prompts = [
torch.randint(0, 256, (16,)), # raw text prompt
(0, torch.randn(4, 384)), # raw modality prompt
None, # empty prompt
[torch.randint(0, 256, (8,)), (0, torch.randn(6, 384))], # interleaved text + modality
]
samples = model.sample_many(prompts, max_length = 2048, cfg_scale = 3.)
the samples come back as a list of ModalitySamples, one per prompt. the kv cache is always used
(the cache_kv = True code path of sample_one), and each sample keeps its own max length budget,
so one sample finishing early does not hold up the rest of the batch.
To pretrain on language first, just pass in your text as type Int['batch seq']
import torch
from transfusion_pytorch import Transfusion
model = Transfusion(
num_text_tokens = 256,
dim_latent = 384,
transformer = dict(
dim = 512,
depth = 8,
)
).cuda()
text = torch.randint(0, 256, (2, 1024)).cuda()
loss = model(text)
loss.backward()
# after much training
sampled = model.generate_text_only(text[:, :1], 1024)
Examples
To run any of the examples train_{example_name}.py in the project root, simply install dependencies first as so
$ pip install .[examples]
If you run into some weird error with safetensors, run this too
$ pip install -U diffusers transformers accelerate scipy ftfy safetensors
Forcing a modality at the start of generation
To force the first generated thing to be a specific modality - skipping text decoding at the start -
pass force_modality_at_start to sample / sample_many with the modality type, or a
(modality_type, shape) tuple to also force its shape. for example, prompt with some text, then a
video (modality 0, channel first 3d latent), and force a 1d action of chunk length 32 (modality 1):
import torch
from transfusion_pytorch import Transfusion
model = Transfusion(
num_text_tokens = 256,
dim_latent = (16, 8),
modality_default_shape = ((4, 4, 4), (16,)), # (video, action)
channel_first_latent = (True, False),
transformer = dict(
dim = 512,
depth = 8
)
)
text = torch.randint(0, 256, (16,)) # some text prompt
video = torch.randn(16, 4, 4, 4) # channel first video latent - (latent dim, frames, height, width)
samples = model.sample_many(
[[text, (0, video)]], # one prompt: text, then video
force_modality_at_start = (1, (32,)), # force a 1d action of chunk length 32
cfg_scale = 3.,
)
action = samples[0][3][1] # the forced action, of shape (32, 8)
the batch can mix prompts of different text and video lengths - each sample still gets its forced action in the same shape.
forcing only the modality type (force_modality_at_start = 1) falls back to the modality's default
shape (or fixed_modality_shape).
Citations
@inproceedings{Zhou2024TransfusionPT,
title = {Transfusion: Predict the Next Token and Diffuse Images with One Multi-Modal Model},
author = {Chunting Zhou and Lili Yu and Arun Babu and Kushal Tirumala and Michihiro Yasunaga and Leonid Shamis and Jacob Kahn and Xuezhe Ma and Luke Zettlemoyer and Omer Levy},
year = {2024},
url = {https://api.semanticscholar.org/CorpusID:271909855}
}
@misc{Rubin2024,
author = {Ohad Rubin},
url = {https://medium.com/@ohadrubin/exploring-weight-decay-in-layer-normalization-challenges-and-a-reparameterization-solution-ad4d12c24950}
}
@article{Nguyen2024MinPS,
title = {Min P Sampling: Balancing Creativity and Coherence at High Temperature},
author = {Minh Nguyen and Andrew Baker and Andreas Kirsch and Clement Neo},
journal = {ArXiv},
year = {2024},
volume = {abs/2407.01082},
url = {https://api.semanticscholar.org/CorpusID:270870613}
}
@article{Bao2022AllAW,
title = {All are Worth Words: A ViT Backbone for Diffusion Models},
author = {Fan Bao and Shen Nie and Kaiwen Xue and Yue Cao and Chongxuan Li and Hang Su and Jun Zhu},
journal = {2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
year = {2022},
pages = {22669-22679},
url = {https://api.semanticscholar.org/CorpusID:253581703}
}
@inproceedings{Zhao2024MonoFormerOT,
title = {MonoFormer: One Transformer for Both Diffusion and Autoregression},
author = {Chuyang Zhao and Yuxing Song and Wenhao Wang and Haocheng Feng and Errui Ding and Yifan Sun and Xinyan Xiao and Jingdong Wang},
year = {2024},
url = {https://api.semanticscholar.org/CorpusID:272832492}
}
@article{Yang2024ConsistencyFM,
title = {Consistency Flow Matching: Defining Straight Flows with Velocity Consistency},
author = {Ling Yang and Zixiang Zhang and Zhilong Zhang and Xingchao Liu and Minkai Xu and Wentao Zhang and Chenlin Meng and Stefano Ermon and Bin Cui},
journal = {ArXiv},
year = {2024},
volume = {abs/2407.02398},
url = {https://api.semanticscholar.org/CorpusID:270878436}
}
@inproceedings{Zhou2024ValueRL,
title = {Value Residual Learning For Alleviating Attention Concentration In Transformers},
author = {Zhanchao Zhou and Tianyi Wu and Zhiyun Jiang and Zhenzhong Lan},
year = {2024},
url = {https://api.semanticscholar.org/CorpusID:273532030}
}
@inproceedings{Duvvuri2024LASERAW,
title = {LASER: Attention with Exponential Transformation},
author = {Sai Surya Duvvuri and Inderjit S. Dhillon},
year = {2024},
url = {https://api.semanticscholar.org/CorpusID:273849947}
}
@inproceedings{Dong2024HymbaAH,
title = {Hymba: A Hybrid-head Architecture for Small Language Models},
author = {Xin Dong and Y. Fu and Shizhe Diao and Wonmin Byeon and Zijia Chen and Ameya Mahabaleshwarkar and Shih-Yang Liu and Matthijs Van Keirsbilck and Min-Hung Chen and Yoshi Suhara and Yingyan Lin and Jan Kautz and Pavlo Molchanov},
year = {2024},
url = {https://api.semanticscholar.org/CorpusID:274166163}
@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}
}
@article{Zhu2025FracConnectionsFE,
title = {Frac-Connections: Fractional Extension of Hyper-Connections},
author = {Defa Zhu and Hongzhi Huang and Jundong Zhou and Zihao Huang and Yutao Zeng and Banggu Wu and Qiyang Min and Xun Zhou},
journal = {ArXiv},
year = {2025},
volume = {abs/2503.14125},
url = {https://api.semanticscholar.org/CorpusID:277104144}
}
@misc{li2025basicsletdenoisinggenerative,
title = {Back to Basics: Let Denoising Generative Models Denoise},
author = {Tianhong Li and Kaiming He},
year = {2025},
eprint = {2511.13720},
archivePrefix = {arXiv},
primaryClass = {cs.CV},
url = {https://arxiv.org/abs/2511.13720},
}
@misc{chefer2026self,
title = {Self-Supervised Flow Matching for Scalable Multi-Modal Synthesis},
author = {Hila Chefer and Patrick Esser and Dominik Lorenz and Dustin Podell and Vikash Raja and Vinh Tong and Antonio Torralba and Robin Rombach},
year = {2026},
url = {https://bfl.ai/research/self-flow},
note = {Preprint}
}
@misc{kimiteam2026attentionresiduals,
title = {Attention Residuals},
author = {Kimi Team and Guangyu Chen and Yu Zhang and Jianlin Su and Weixin Xu and Siyuan Pan and Yaoyu Wang and Yucheng Wang and Guanduo Chen and Bohong Yin and Yutian Chen and Junjie Yan and Ming Wei and Y. Zhang and Fanqing Meng and Chao Hong and Xiaotong Xie and Shaowei Liu and Enzhe Lu and Yunpeng Tai and Yanru Chen and Xin Men and Haiqing Guo and Y. Charles and Haoyu Lu and Lin Sui and Jinguo Zhu and Zaida Zhou and Weiran He and Weixiao Huang and Xinran Xu and Yuzhi Wang and Guokun Lai and Yulun Du and Yuxin Wu and Zhilin Yang and Xinyu Zhou},
year = {2026},
eprint = {2603.15031},
archivePrefix = {arXiv},
primaryClass = {cs.CL},
url = {https://arxiv.org/abs/2603.15031},
}
1 maintainer
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
Built Distribution
Filter files by name, interpreter, ABI, and platform.
If you're not sure about the file name format, learn more about wheel file names.
Copy a direct link to the current filters
File details
Details for the file transfusion_pytorch-0.19.5.tar.gz.
File metadata
- Download URL: transfusion_pytorch-0.19.5.tar.gz
- Upload date:
- Size: 45.0 kB
- Tags: Source
- Uploaded using Trusted Publishing? No
- Uploaded via:
uv/0.8.17
File hashes
| Algorithm | Hash digest | |
|---|---|---|
| SHA256 |
9b9ec0ed1e081b4c213aaf9bc371a05c5dd68ade73ba54d6cab7658fdd10bc69
|
|
| MD5 |
39a0c194ecafabc429eff3a25a8711e2
|
|
| BLAKE2b-256 |
6f22487ad2291d20dc306fe1da28e88a5303f6459989740cde1acd9795314691
|
File details
Details for the file transfusion_pytorch-0.19.5-py3-none-any.whl.
File metadata
- Download URL: transfusion_pytorch-0.19.5-py3-none-any.whl
- Upload date:
- Size: 45.8 kB
- Tags: Python 3
- Uploaded using Trusted Publishing? No
- Uploaded via:
uv/0.8.17
File hashes
| Algorithm | Hash digest | |
|---|---|---|
| SHA256 |
ed69c3b4d430192e997f23a693b18583050476b051b94bca1167ca041c5fd38a
|
|
| MD5 |
0de42c39ebd760d42bafa8e644fecb9d
|
|
| BLAKE2b-256 |
e0c45ed60ef9f998ed3d8cc5e20905276fd3a2723ebf96ceda828a5e0eb006c3
|
