StonyBrook-CVLab/PixCell-1024

import torch

from diffusers import DiffusionPipeline
from diffusers import AutoencoderKL

device = torch.device('cuda')

# We do not host the weights of the SD3 VAE -- load it from StabilityAI
sd3_vae = AutoencoderKL.from_pretrained("stabilityai/stable-diffusion-3.5-large", subfolder="vae")

pipeline = DiffusionPipeline.from_pretrained(
    "StonyBrook-CVLab/PixCell-1024",
    vae=sd3_vae,
    custom_pipeline="StonyBrook-CVLab/PixCell-pipeline",
    trust_remote_code=True,
    torch_dtype=torch.float16,
)

pipeline.to(device);

import timm
from timm.data import resolve_data_config
from timm.data.transforms_factory import create_transform

timm_kwargs = {
            'img_size': 224, 
            'patch_size': 14, 
            'depth': 24,
            'num_heads': 24,
            'init_values': 1e-5, 
            'embed_dim': 1536,
            'mlp_ratio': 2.66667*2,
            'num_classes': 0, 
            'no_embed_class': True,
            'mlp_layer': timm.layers.SwiGLUPacked, 
            'act_layer': torch.nn.SiLU, 
            'reg_tokens': 8, 
            'dynamic_img_size': True
        }
uni_model = timm.create_model("hf-hub:MahmoodLab/UNI2-h", pretrained=True, **timm_kwargs)
transform = create_transform(**resolve_data_config(uni_model.pretrained_cfg, model=uni_model))
uni_model.eval()
uni_model.to(device);

uncond = pipeline.get_unconditional_embedding(1)
with torch.amp.autocast('cuda'):
    samples = pipeline(uni_embeds=uncond, negative_uni_embeds=None, guidance_scale=1.0)

# Load image
import numpy as np
import einops
from PIL import Image
from huggingface_hub import hf_hub_download

# This is an example image we provide
path = hf_hub_download(repo_id="StonyBrook-CVLab/PixCell-1024", filename="test_image.png")
image = Image.open(path).convert("RGB")


# Rearrange 1024x1024 image into 16 256x256 patches
uni_patches = np.array(image)
uni_patches = einops.rearrange(uni_patches, '(d1 h) (d2 w) c -> (d1 d2) h w c', d1=4, d2=4)
uni_input = torch.stack([transform(Image.fromarray(item)) for item in uni_patches])

# Extract UNI embeddings
with torch.inference_mode():
    uni_emb = uni_model(uni_input.to(device))

# reshape UNI to (bs, 16, D)
uni_emb = uni_emb.unsqueeze(0)
print("Extracted UNI:", uni_emb.shape)

# Get unconditional embedding for classifier-free guidance
uncond = pipeline.get_unconditional_embedding(uni_emb.shape[0])
# Generate new samples
with torch.amp.autocast('cuda'):
    samples = pipeline(uni_embeds=uni_emb, negative_uni_embeds=uncond, guidance_scale=1.5).images