app.py

import os
import sys
import git
import gradio as gr
import torch
from huggingface_hub import hf_hub_download
from diffusers import DiffusionPipeline

# Check if the cog-sdxl repository exists, otherwise clone it
repo_url = "https://github.com/replicate/cog-sdxl.git"
repo_dir = "./cog-sdxl"

if not os.path.exists(repo_dir):
    print("Cloning cog-sdxl repository...")
    git.Repo.clone_from(repo_url, repo_dir)

# Add cog_sdxl directory to sys.path for imports
sys.path.append(os.path.abspath(repo_dir))

# Now import the required module after adding cog_sdxl to sys.path
from cog_sdxl.dataset_and_utils import TokenEmbeddingsHandler

# Load the model pipeline
pipe = DiffusionPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0",
    torch_dtype=torch.float16,
    variant="fp16",
).to("cuda")

# Load the LoRA weights
pipe.load_lora_weights("fofr/sdxl-emoji", weight_name="lora.safetensors")

# Load token embeddings
embedding_path = hf_hub_download(repo_id="fofr/sdxl-emoji", filename="embeddings.pti", repo_type="model")
text_encoders = [pipe.text_encoder, pipe.text_encoder_2]
tokenizers = [pipe.tokenizer, pipe.tokenizer_2]
embhandler = TokenEmbeddingsHandler(text_encoders, tokenizers)
embhandler.load_embeddings(embedding_path)

def generate_emoji(prompt):
    """Generate an emoji image based on the user's prompt."""
    prompt = f"A <s0><s1> emoji of {prompt}"
    images = pipe(
        prompt,
        cross_attention_kwargs={"scale": 0.8},
    ).images
    return images[0]

# Gradio UI definition
iface = gr.Interface(
    fn=generate_emoji,
    inputs=gr.Textbox(label="Enter description for emoji"),
    outputs=gr.Image(label="Generated Emoji"),
    title="SDXL Emoji Generator",
    description="Generate a custom emoji using SDXL model with LoRA fine-tuning."
)

if __name__ == "__main__":
    iface.launch()