app.py

import os
from huggingface_hub import login, snapshot_download
from transformers import BlipProcessor, BlipForConditionalGeneration
from PIL import Image
from dotenv import load_dotenv
import gradio as gr
from diffusers import FluxPipeline
import torch
import spaces  # Hugging Face Spaces module

# -----------------------
# Pre-cache models at startup
# -----------------------
snapshot_download("Salesforce/blip-image-captioning-large", timeout=120)
snapshot_download("noamrot/FuseCap", timeout=120)
snapshot_download("black-forest-labs/FLUX.1-dev", timeout=300)

# -----------------------
# Authentication
# -----------------------
load_dotenv()
HUGGINGFACE_TOKEN = os.getenv("HUGGINGFACE_TOKEN")
if HUGGINGFACE_TOKEN:
    login(token=HUGGINGFACE_TOKEN)

# -----------------------
# Load models
# -----------------------
device = "cuda" if torch.cuda.is_available() else "cpu"

processor = BlipProcessor.from_pretrained("Salesforce/blip-image-captioning-large", timeout=120)
model = BlipForConditionalGeneration.from_pretrained("Salesforce/blip-image-captioning-large", timeout=120).to(device)

processor1 = BlipProcessor.from_pretrained("noamrot/FuseCap", timeout=120)
model2 = BlipForConditionalGeneration.from_pretrained("noamrot/FuseCap", timeout=120).to(device)

pipe = FluxPipeline.from_pretrained("black-forest-labs/FLUX.1-dev", torch_dtype=torch.bfloat16, timeout=300).to(device)

# -----------------------
# Options
# -----------------------
fabrics = ['cotton', 'silk', 'denim', 'linen', 'polyester', 'wool', 'velvet']
patterns = ['striped', 'floral', 'geometric', 'abstract', 'solid', 'polka dots']
textile_designs = ['woven texture', 'embroidery', 'printed fabric', 'hand-dyed', 'quilting']

# -----------------------
# Inference Function
# -----------------------
@spaces.GPU(duration=150)
def generate_caption_and_image(image, f, p, d):
    if image and f and p and d:
        img = image.convert("RGB")

        # Caption with FuseCap
        inputs = processor(img, "a picture of ", return_tensors="pt").to(device)
        out = model2.generate(**inputs, num_beams=3)
        caption2 = processor1.decode(out[0], skip_special_tokens=True)

        # Caption with BLIP
        inputs = processor(image, return_tensors="pt", padding=True, truncation=True, max_length=250)
        inputs = {k: v.to(device) for k, v in inputs.items()}
        out = model.generate(**inputs)
        caption1 = processor.decode(out[0], skip_special_tokens=True)

        # Compose prompt
        prompt = (
            f"Design a high-quality, stylish clothing item that combines the essence of {caption1} and {caption2}. "
            f"Use luxurious {f} fabric with intricate {d} design elements. "
            f"Incorporate {p} patterns to elevate the garment's aesthetic. "
            "Ensure sophistication, innovation, and timeless elegance."
        )

        # Generate image
        result = pipe(
            prompt,
            height=1024,
            width=1024,
            guidance_scale=3.5,
            num_inference_steps=50,
            max_sequence_length=512,
            generator=torch.Generator('cpu').manual_seed(0)
        ).images[0]

        return result
    return None

# -----------------------
# Gradio UI
# -----------------------
iface = gr.Interface(
    fn=generate_caption_and_image,
    inputs=[
        gr.Image(type="pil", label="Upload Image"),
        gr.Radio(fabrics, label="Select Fabric"),
        gr.Radio(patterns, label="Select Pattern"),
        gr.Radio(textile_designs, label="Select Textile Design")
    ],
    outputs=gr.Image(label="Generated Design"),
    live=True
)

iface.launch()