app.py

import os
import pickle
import numpy as np
import gradio as gr
import torch
from datasets import load_dataset
from PIL import Image

from transformers import AutoImageProcessor, AutoModel
from transformers import CLIPProcessor, CLIPModel

# Faster generation model (Turbo)
from diffusers import AutoPipelineForText2Image


# ==========================================================
# CONFIG
# ==========================================================
DATASET_ID = "MAY199/synthetic-sofa-images"
PICKLE_PATH = "sofa_embeddings_for_app.pkl"

# Fast model for quick generation
SD_MODEL_ID = "stabilityai/sd-turbo"
CLIP_MODEL_ID = "openai/clip-vit-base-patch32"

# Make sure it generates a SOFA (avoid stripes/patterns)
NEGATIVE_PROMPT_BASE = (
    "low quality, blurry, distorted, watermark, text, logo, deformed, bad proportions, "
    "cartoon, painting, sketch, bad anatomy, extra legs, extra cushions, "
    "abstract, pattern, texture, stripes, lines, wallpaper, fabric sample, swatch, "
    "gradient, color blocks, flat image"
)

STYLE_BANK = [
    "modern minimalist",
    "scandinavian",
    "mid century modern",
    "contemporary",
    "japandi",
    "industrial modern",
    "luxury modern",
    "coastal minimal",
]

MATERIAL_BANK = [
    "linen fabric",
    "velvet fabric",
    "boucle fabric",
    "leather",
    "microfiber fabric",
]

SHOT_BANK = [
    "studio product photo, isolated object, seamless white background, softbox lighting, sharp focus, ultra realistic",
    "clean catalog photo, product cutout, white seamless background, softbox lighting, sharp focus, ultra realistic",
]


# ==========================================================
# LOAD DATASET (HF)
# ==========================================================
ds = load_dataset(DATASET_ID)
train_ds = ds["train"]


# ==========================================================
# LOAD EMBEDDINGS (ViT) for FAST retrieval
# ==========================================================
with open(PICKLE_PATH, "rb") as f:
    data = pickle.load(f)

vit_model_id = data["model_id"]
emb_matrix = data["embeddings"].astype(np.float32)
image_indices = data["image_indices"]

if "vit" not in vit_model_id.lower():
    raise ValueError(f"This app expects a ViT model_id for retrieval, got: {vit_model_id}")


# ==========================================================
# DEVICE
# ==========================================================
device = "cuda" if torch.cuda.is_available() else "cpu"
print("Device:", device)


# ==========================================================
# LOAD ViT (retrieval encoder)
# ==========================================================
vit_processor = AutoImageProcessor.from_pretrained(vit_model_id, use_fast=True)
vit_model = AutoModel.from_pretrained(vit_model_id).to(device)
vit_model.eval()

def l2_normalize(x: np.ndarray) -> np.ndarray:
    return x / (np.linalg.norm(x) + 1e-12)

@torch.no_grad()
def embed_image_vit(img: Image.Image) -> np.ndarray:
    inputs = vit_processor(images=img, return_tensors="pt")
    inputs = {k: v.to(device) for k, v in inputs.items()}
    outputs = vit_model(**inputs)
    feats = outputs.last_hidden_state.mean(dim=1)
    feats = feats / feats.norm(dim=-1, keepdim=True)
    return feats.squeeze(0).float().cpu().numpy()


# ==========================================================
# LOAD CLIP (style selection only)
# ==========================================================
clip_processor = CLIPProcessor.from_pretrained(CLIP_MODEL_ID)
clip_model = CLIPModel.from_pretrained(CLIP_MODEL_ID).to(device)
clip_model.eval()

@torch.no_grad()
def choose_style_with_clip(input_img: Image.Image, style_prompts: list[str]) -> str:
    inputs = clip_processor(
        text=style_prompts,
        images=input_img,
        return_tensors="pt",
        padding=True
    )
    inputs = {k: v.to(device) for k, v in inputs.items()}
    outputs = clip_model(**inputs)
    scores = outputs.logits_per_image.squeeze(0)
    best_idx = int(torch.argmax(scores).item())
    return style_prompts[best_idx]


# ==========================================================
# COLOR EXTRACTION (fast and simple): use mean RGB -> nearest names
# ==========================================================
NAMED_COLORS_RGB = {
    "white": (245, 245, 245),
    "black": (20, 20, 20),
    "gray": (128, 128, 128),
    "beige": (215, 198, 170),
    "brown": (120, 75, 45),
    "tan": (180, 140, 95),
    "cream": (240, 230, 205),

    "red": (210, 40, 40),
    "pink": (235, 145, 175),
    "magenta": (200, 60, 160),
    "purple": (120, 70, 170),

    "orange": (230, 140, 50),
    "yellow": (235, 210, 60),

    "green": (60, 160, 90),
    "olive": (120, 140, 60),
    "teal": (50, 150, 150),
    "cyan": (70, 190, 210),

    "blue": (60, 110, 205),
    "navy": (35, 55, 110),
}

def nearest_named_colors(img: Image.Image, top_n: int = 2):
    im = img.convert("RGB").resize((64, 64))
    arr = np.array(im).reshape(-1, 3).astype(np.float32)
    mean_rgb = arr.mean(axis=0)

    # pick nearest by euclidean RGB distance (fast)
    dists = []
    for name, rgb in NAMED_COLORS_RGB.items():
        d = float(((mean_rgb - np.array(rgb)) ** 2).sum())
        dists.append((d, name))
    dists.sort()
    names = [name for _, name in dists[:top_n]]
    return names


# ==========================================================
# SD Turbo txt2img (lazy-load)
# ==========================================================
sd_pipe = None

def get_sd_pipe():
    global sd_pipe
    if sd_pipe is not None:
        return sd_pipe

    dtype = torch.float16 if device == "cuda" else torch.float32
    p = AutoPipelineForText2Image.from_pretrained(
        SD_MODEL_ID,
        torch_dtype=dtype,
        safety_checker=None,
    ).to(device)

    # Turbo-friendly defaults
    if device == "cuda":
        try:
            p.enable_attention_slicing()
        except Exception:
            pass

    sd_pipe = p
    return sd_pipe


# ==========================================================
# FAST RETRIEVAL (dataset) using ViT embeddings
# ==========================================================
def recommend_fast(img: Image.Image):
    if img is None:
        raise gr.Error("Please upload an image first.")

    q = l2_normalize(embed_image_vit(img)).astype(np.float32)
    sims = emb_matrix @ q
    top_idx = np.argsort(-sims)

    results = []
    for j in top_idx:
        if sims[j] > 0.999:
            continue
        ds_idx = image_indices[j]
        results.append(train_ds[int(ds_idx)]["image"])
        if len(results) == 3:
            break

    while len(results) < 3:
        ds_idx = image_indices[int(top_idx[len(results)])]
        results.append(train_ds[int(ds_idx)]["image"])

    return results[0], results[1], results[2]


# ==========================================================
# AI GENERATION (FAST + "SOFA GUARANTEE")
# - strong sofa geometry words
# - strong negative for patterns/stripes
# - Turbo few steps for speed
# ==========================================================
def build_style_prompts():
    prompts = []
    for style in STYLE_BANK:
        for material in MATERIAL_BANK:
            for shot in SHOT_BANK:
                prompts.append(f"{style}, {material}, {shot}")
    return prompts

STYLE_PROMPTS = build_style_prompts()

def generate_new_sofa_from_input_only(img: Image.Image, seed: int = 0):
    if img is None:
        raise gr.Error("Please upload an image first.")

    p = get_sd_pipe()

    # colors from input only (fast)
    color_names = nearest_named_colors(img, top_n=2)
    color_phrase = " and ".join(color_names)

    # style from CLIP (input-only)
    chosen_style = choose_style_with_clip(img, STYLE_PROMPTS)

    # Strong geometry anchors to avoid "stripes"
    prompt = (
        "a realistic three-dimensional sofa, seating furniture, "
        "with visible cushions, armrests, and legs, "
        f"{color_phrase} color, {chosen_style}, "
        "single sofa only, isolated object, product cutout, "
        "no room, no living room, no interior, no background scene"
    )

    negative_prompt = NEGATIVE_PROMPT_BASE

    generator = None
    if seed and int(seed) > 0:
        generator = torch.Generator(device=device).manual_seed(int(seed))

    # SD Turbo fast settings
    out = p(
        prompt=prompt,
        negative_prompt=negative_prompt,
        num_inference_steps=4,
        guidance_scale=1.5,
        height=512,
        width=512,
        generator=generator,
    )

    img_out = out.images[0]
    return img_out, prompt, str(color_names)


# ==========================================================
# UI
# ==========================================================
with gr.Blocks() as app:
    gr.Markdown("# Sofa Recommendation System + AI Generation")
    gr.Markdown(
        "Submit (Fast) returns 3 similar sofas from the dataset using ViT embeddings. "
        "Generate AI creates a NEW sofa image influenced ONLY by the input image via auto prompt."
    )

    inp = gr.Image(type="pil", label="Upload a living room or sofa image")

    gr.Markdown("## Quick Starters (1-click examples)")
    gr.Examples(
        examples=[
            "examples/starter1.jpeg",
            "examples/starter2.jpeg",
            "examples/starter3.jpeg",
        ],
        inputs=inp,
        label="Quick Starters",
    )

    with gr.Row():
        btn_submit = gr.Button("Submit (Fast) - Dataset Retrieval")
        btn_ai = gr.Button("Generate AI (New Sofa) - Input Only")

    with gr.Row():
        seed_inp = gr.Slider(minimum=0, maximum=999999, step=1, value=0, label="Seed (0 = random)")
        palette_box = gr.Textbox(label="Detected colors (names from INPUT only)", lines=1)

    chosen_prompt_box = gr.Textbox(label="Auto prompt used for generation (built from INPUT only)", lines=4)

    with gr.Row():
        out1 = gr.Image(label="Recommendation 1 (dataset)")
        out2 = gr.Image(label="Recommendation 2 (dataset)")
        out3 = gr.Image(label="Recommendation 3 (dataset)")

    out_ai = gr.Image(label="AI Generated Sofa (new image, no dataset image used)")

    btn_submit.click(fn=recommend_fast, inputs=inp, outputs=[out1, out2, out3])
    btn_ai.click(fn=generate_new_sofa_from_input_only, inputs=[inp, seed_inp], outputs=[out_ai, chosen_prompt_box, palette_box])

app.launch()