from PIL import Image
import torch
from muse import PipelineMuse, MaskGiTUViT
from datasets import Dataset, Features
from datasets import Image as ImageFeature
from datasets import Value, load_dataset
device = "cuda" if torch.cuda.is_available() else "cpu"
pipe = PipelineMuse.from_pretrained(
transformer_path="valhalla/research-run",
text_encoder_path="openMUSE/clip-vit-large-patch14-text-enc",
vae_path="openMUSE/vqgan-f16-8192-laion",
).to(device)
pipe.transformer = MaskGiTUViT.from_pretrained("valhalla/muse-research-run", subfolder="ema_model").to(device)
pipe.tokenizer.pad_token_id = 49407
if device == "cuda":
pipe.transformer.enable_xformers_memory_efficient_attention()
pipe.text_encoder.to(torch.float16)
pipe.transformer.to(torch.float16)
import PIL
def main():
print("Loading dataset...")
parti_prompts = load_dataset("nateraw/parti-prompts", split="train")
print("Loading pipeline...")
seed = 0
device = "cuda"
torch.manual_seed(0)
ckpt_id = "openMUSE/muse-256"
scale = 10
print("Running inference...")
main_dict = {}
for i in range(len(parti_prompts)):
sample = parti_prompts[i]
prompt = sample["Prompt"]
image = pipe(
prompt,
timesteps=16,
negative_text=None,
guidance_scale=scale,
temperature=(2, 0),
orig_size=(256, 256),
crop_coords=(0, 0),
aesthetic_score=6,
use_fp16=device == "cuda",
transformer_seq_len=256,
use_tqdm=False,
)[0]
image = image.resize((256, 256), resample=PIL.Image.Resampling.LANCZOS)
img_path = f"/home/patrick/muse_images/muse_256_{i}.png"
image.save(img_path)
main_dict.update(
{
prompt: {
"img_path": img_path,
"Category": sample["Category"],
"Challenge": sample["Challenge"],
"Note": sample["Note"],
"model_name": ckpt_id,
"seed": seed,
}
}
)
def generation_fn():
for prompt in main_dict:
prompt_entry = main_dict[prompt]
yield {
"Prompt": prompt,
"Category": prompt_entry["Category"],
"Challenge": prompt_entry["Challenge"],
"Note": prompt_entry["Note"],
"images": {"path": prompt_entry["img_path"]},
"model_name": prompt_entry["model_name"],
"seed": prompt_entry["seed"],
}
print("Preparing HF dataset...")
ds = Dataset.from_generator(
generation_fn,
features=Features(
Prompt=Value("string"),
Category=Value("string"),
Challenge=Value("string"),
Note=Value("string"),
images=ImageFeature(),
model_name=Value("string"),
seed=Value("int64"),
),
)
ds_id = "diffusers-parti-prompts/muse256"
ds.push_to_hub(ds_id)
if __name__ == "__main__":
main()
