app.py

import os, pathlib, io
import numpy as np, cv2, torch, trimesh
from PIL import Image
import gradio as gr

from segment_anything import sam_model_registry, SamPredictor
from diffusers import StableDiffusionControlNetPipeline

# SHAP-E main branch (import minimal, potresti dover adattare in futuro)
try:
    from shap_e.diffusion.sample import build_model
except ImportError:
    build_model = None

# — checkpoint directory —
CKPT_DIR = pathlib.Path("checkpoints")
CKPT_DIR.mkdir(exist_ok=True)
SAM_PATH = CKPT_DIR / "sam_vit_b.pth"
CN_PATH  = CKPT_DIR / "controlnet_scribble.bin"

def fetch(url, dst):
    if dst.exists(): return
    os.system(f"wget -q {url} -O {dst}")

# scarica i checkpoint se assenti
fetch(
    "https://github.com/facebookresearch/segment-anything/releases/download/v0.1.0/sam_vit_b_01ec64.pth",
    SAM_PATH,
)
fetch(
    "https://huggingface.co/lllyasviel/control_v11p_sd15_scribble/resolve/main/pytorch_model.bin",
    CN_PATH,
)

device = torch.device("cpu")

# SAM
sam   = sam_model_registry["vit_b"](checkpoint=str(SAM_PATH)).to(device)
predictor = SamPredictor(sam)

# Stable Diffusion + ControlNet
controlnet = StableDiffusionControlNetPipeline.from_pretrained(
    "lllyasviel/sd-controlnet-scribble", torch_dtype=torch.float32
).controlnet.to(device)
pipe = StableDiffusionControlNetPipeline.from_pretrained(
    "runwayml/stable-diffusion-v1-5",
    controlnet=controlnet,
    torch_dtype=torch.float32
).to(device)
pipe.enable_attention_slicing()

# SHAP-E (se presente)
if build_model:
    shap_e_model = build_model("diffusion", device=device)
    shap_e_model.eval()
else:
    shap_e_model = None

# — helper pipeline —
def get_silhouette(img: Image.Image):
    arr = np.array(img.convert("RGB"))
    predictor.set_image(arr)
    mask, _, _ = predictor.predict(
        point_coords=None,
        point_labels=None,
        box=None,
        multimask_output=False
    )
    return (mask[0].astype(np.uint8) * 255)

def generate_concept(mask: np.ndarray, prompt: str):
    mask_rgb = cv2.cvtColor(mask, cv2.COLOR_GRAY2RGB)
    out = pipe(
        prompt=prompt,
        image=Image.fromarray(mask_rgb),
        num_inference_steps=30,
        guidance_scale=7.5
    )
    return out.images[0]

def generate_mesh(concept_img: Image.Image):
    if shap_e_model is None:
        # fallback stub: restituisco un cubo STL minimo
        cube = trimesh.creation.box(extents=(1,1,1))
        return cube
    buf = io.BytesIO()
    concept_img.save(buf, format="PNG")
    buf.seek(0)
    # decodifica SHAP-E (API stabile se build_model esiste)
    mesh = shap_e_model.sample_latents_and_decode(buf)
    return trimesh.Trimesh(vertices=mesh.vertices, faces=mesh.faces)

def export_stl(mesh: trimesh.Trimesh):
    out_path = "output_eol.stl"
    mesh.export(out_path)
    return out_path

def eol_pipeline(image, prompt):
    mask    = get_silhouette(image)
    concept = generate_concept(mask, prompt)
    mesh    = generate_mesh(concept)
    stl     = export_stl(mesh)
    return concept, stl

# — Gradio UI —
demo = gr.Interface(
    fn=eol_pipeline,
    inputs=[
        gr.Image(type="pil", label="Foto componente EoL"),
        gr.Textbox(lines=2, placeholder="Prompt creativo…")
    ],
    outputs=[
        gr.Image(type="pil", label="Concept 2D"),
        gr.File(label="Scarica STL")
    ],
    title="EoL Component Generative Design (CPU-only)",
    description="Upload foto componente EoL ➜ concept 2D ➜ mesh 3D STL"
)

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