app.py

import torch
from diffusers import StableDiffusion3Pipeline
from huggingface_hub import login
import os
import gradio as gr
from diffusers import BitsAndBytesConfig
from diffusers import SD3Transformer2DModel

# Retrieve the token from the environment variable
token = os.getenv("HF_TOKEN")  # Hugging Face token from the secret
if token:
    login(token=token)  # Log in with the retrieved token
else:
    raise ValueError("Hugging Face token not found. Please set it as a repository secret in the Space settings.")

# Define quantization configuration (4-bit quantization)
quant_config = BitsAndBytesConfig(
    load_in_4bit=True,  # Enable 4-bit quantization
    bnb_4bit_quant_type="nf4",  # Choose the quantization type (nf4 is often used for high-quality quantization)
    bnb_4bit_compute_dtype=torch.bfloat16  # Use bfloat16 for computation (works well with CPUs)
)

# Load the Stable Diffusion 3.5 model with quantization
model_id = "stabilityai/stable-diffusion-3.5-large"
model = SD3Transformer2DModel.from_pretrained(
    model_id,
    subfolder="transformer",
    quantization_config=quant_config,
    torch_dtype=torch.bfloat16  # Ensure the model uses bfloat16 dtype for computation
)

# Load the pipeline with the quantized model
pipe = StableDiffusion3Pipeline.from_pretrained(
    model_id,
    transformer=model,
    torch_dtype=torch.bfloat16  # Ensuring the pipeline uses bfloat16
)

pipe.to("cpu")  # Ensuring it runs on CPU

# Define the path to the LoRA model
lora_model_path = "./lora_model.pth"  # Assuming the file is saved locally

# Custom method to load and apply LoRA weights to the Stable Diffusion pipeline
def load_lora_model(pipe, lora_model_path):
    # Load the LoRA weights
    lora_weights = torch.load(lora_model_path, map_location="cpu")
    
    # Apply weights to the UNet submodule
    for name, param in pipe.unet.named_parameters():  # Accessing unet parameters
        if name in lora_weights:
            param.data += lora_weights[name]

    return pipe

# Load and apply the LoRA model weights
pipe = load_lora_model(pipe, lora_model_path)

# Function to generate an image from a text prompt
def generate_image(prompt):
    image = pipe(prompt).images[0]
    return image

# Gradio interface
iface = gr.Interface(fn=generate_image, inputs="text", outputs="image")
iface.launch()