|
|
--- |
|
|
{} |
|
|
--- |
|
|
``` |
|
|
import os |
|
|
import torch |
|
|
if not os.path.exists("seed2"): |
|
|
os.system("git clone https://huggingface.co/ontocord/seed2") |
|
|
from diffusers import StableUnCLIPImg2ImgPipeline |
|
|
from seed2.seed2_tokenizer import Seed2Tokenizer |
|
|
tokenizer = None |
|
|
pipe = None |
|
|
torch.cuda.empty_cache() |
|
|
pipe = StableUnCLIPImg2ImgPipeline.from_pretrained('stabilityai/stable-diffusion-2-1-unclip', torch_dtype=torch.float16) |
|
|
pipe = pipe.to('cuda') |
|
|
tokenizer = Seed2Tokenizer.from_pretrained("seed2", torch_dtype=torch.float16).to('cuda') |
|
|
if not os.path.exists("cat.jpg"): |
|
|
os.system("wget https://images.unsplash.com/photo-1574158622682-e40e69881006?w=300 -O cat.jpg") |
|
|
tokens = tokenizer.encode_image("cat.jpg") |
|
|
print (tokens) |
|
|
image_embeds = tokenizer.model.get_codebook_entry(tokens) |
|
|
print (image_embeds) |
|
|
import time |
|
|
t = time.time() |
|
|
image = tokenizer.decode(pipe, tokens)[0] # this is using a fixed latent initalized by the model, so a call with the same tokens will produce the same image |
|
|
print (time.time()-t) |
|
|
image |
|
|
``` |
|
|
|
