utils.py

import os
os.environ["HF_HOME"] = "/data/huggingface"
os.environ["TRANSFORMERS_CACHE"] = "/data/huggingface"
os.makedirs("/data/huggingface/hub", exist_ok=True)

import torch
from diffusers import StableDiffusionImg2ImgPipeline
from PIL import Image


# --- Place any download or path setup here ---
MODEL_ID ="runwayml/stable-diffusion-v1-5"  # Can swap for custom path if using IP-Adapter
ADAPTER_PATH = "/workspace/.cache/huggingface/ip_adapter/ip-adapter_sd15.bin"
ADAPTER_DIR = "/workspace/.cache/huggingface/ip_adapter"
DEVICE = "cpu"
MODEL_CACHE = "/workspace/.cache/huggingface"

# (Optional) Download IP-Adapter weights and patch pipeline if desired

# Load the model ONCE at startup, not per request!
pipe = StableDiffusionImg2ImgPipeline.from_pretrained(
    MODEL_ID,
    torch_dtype=torch.float32,
    cache_dir=MODEL_CACHE,
    # safety_checker=None,  # Disable for demo/testing; enable in prod
).to(DEVICE)

pipe.load_ip_adapter(
    pretrained_model_name_or_path_or_dict="h94/IP-Adapter",
    subfolder="models",
    weight_name="ip-adapter_sd15.bin"
)

# pipe.load_ip_adapter(
#     pretrained_model_name_or_path_or_dict=ADAPTER_DIR,
#     subfolder=".",  # The weights file is directly in ADAPTER_DIR
#     weight_name="ip-adapter_sd15.bin"
#     # Optionally: subfolder="models" if using the repo, not a direct path
# )

def generate_sticker(input_image, prompt):
    """
    Given a user image and a prompt, generates a sticker/emoji-style portrait.
    """
    # Load the model (download if not present)
    # pipe = StableDiffusionImg2ImgPipeline.from_pretrained(
    #     MODEL_ID,
    #     torch_dtype=torch.float32,
    #     cache_dir=MODEL_CACHE,
    #     safety_checker=None,  # Disable for demo/testing
    # ).to(DEVICE)

    # Preprocess the image (resize, etc)
    init_image = input_image.convert("RGB").resize((512, 512))

    # Run inference (low strength for identity preservation)
    result = pipe(
        prompt=prompt,
        image=init_image,
        strength=0.65,
        guidance_scale=7.5,
        num_inference_steps=30
    )
    # Return the generated image (as PIL)
    return result.images[0]