Model Card for muse-large-imagenet
<!-- Provide a quick summary of what the model is/does. -->
A MUSE model trained on imagenet as part of the open-muse effort.
Model Details
Model Description
A muse-large model with t5-v1.1-large as the text encoder and a transformer model with the following configuration:
| num_layers | num_heads | hidden_size | intermediate_size |
|---|---|---|---|
| 24 | 16 | 1024 | 4096 |
The model is trained on the imagenet dataset for text2image generation using imagenet class ids as the text.
Model Sources
<!-- Provide the basic links for the model. -->
- Repository: open-muse
- Paper: Muse: Text-To-Image Generation via Masked Generative Transformers
How to Get Started with the Model
First clone and install the open-muse repo per the readme's instructions. And use the following code snippet:
from muse import PipelineMuse
pipe = PipelineMuse.from_pretrained("openMUSE/muse-large-imagenet").to("cuda")
prompt = "goose" # or any other imagenet class id
images = pipe(prompt, num_images_per_prompt=4, timesteps=8, guidance_scale=2, use_maskgit_generate=True)
Some non cherry picked sampler generated by the model:
for more samples, checkout this page.
Training Details
Training Data
The model is trained on the imagenet dataset using the class ids as the text.
Training Procedure
Please refer to the repo and wandb logs for details about the training.
Evaluation
TODO: