app.py

# app.py — ZeroGPU対応版
import gradio as gr
import spaces
import torch
import numpy as np
from PIL import Image
import os
import subprocess
import traceback
import base64
import io
from pathlib import Path

# FastAPI関連（ハイブリッド構成のため維持）
from fastapi import FastAPI, UploadFile, File, Form, HTTPException

# グローバル変数としてパイプラインを定義（初期値はNone）
pipe = None
face_app = None
upsampler = None
UPSCALE_OK = False

# 0. Cache dir & helpers (起動時に実行)
PERSIST_BASE = Path("/data")
CACHE_ROOT = (PERSIST_BASE / "instantid_cache" if PERSIST_BASE.exists() and os.access(PERSIST_BASE, os.W_OK)
              else Path.home() / ".cache" / "instantid_cache")
MODELS_DIR, LORA_DIR, EMB_DIR, UPSCALE_DIR = CACHE_ROOT/"models", CACHE_ROOT/"models"/"Lora", CACHE_ROOT/"embeddings", CACHE_ROOT/"realesrgan"

for p in (MODELS_DIR, LORA_DIR, EMB_DIR, UPSCALE_DIR):
    p.mkdir(parents=True, exist_ok=True)

def dl(url: str, dst: Path, attempts: int = 2):
    if dst.exists(): return
    for i in range(1, attempts + 1):
        print(f"⬇ Downloading {dst.name} (try {i}/{attempts})")
        if subprocess.call(["wget", "-q", "-O", str(dst), url]) == 0: return
    raise RuntimeError(f"download failed → {url}")

# 1. Asset download (起動時に実行)
print("— Starting asset download check —")
BASE_CKPT = MODELS_DIR / "beautiful_realistic_asians_v7_fp16.safetensors"
dl("https://civitai.com/api/download/models/177164?type=Model&format=SafeTensor&size=pruned&fp=fp16", BASE_CKPT)
IP_BIN_FILE = LORA_DIR / "ip-adapter-plus-face_sd15.bin"
dl("https://huggingface.co/h94/IP-Adapter/resolve/main/models/ip-adapter-plus-face_sd15.bin", IP_BIN_FILE)
LORA_FILE = LORA_DIR / "ip-adapter-faceid-plusv2_sd15_lora.safetensors"
dl("https://huggingface.co/h94/IP-Adapter-FaceID/resolve/main/ip-adapter-faceid-plusv2_sd15_lora.safetensors", LORA_FILE)
print("— Asset download check finished —")


# 2. パイプライン初期化関数 (GPU確保後に呼び出される)
def initialize_pipelines():
    global pipe, face_app, upsampler, UPSCALE_OK
    
    # torch/diffusers/onnxruntimeなどのインポートを関数内に移動
    from diffusers import StableDiffusionPipeline, ControlNetModel, DPMSolverMultistepScheduler, AutoencoderKL
    from insightface.app import FaceAnalysis

    print("--- Initializing Pipelines (GPU is now available) ---")
    
    device = torch.device("cuda") # ZeroGPUではGPUが保証されている
    dtype = torch.float16

    # FaceAnalysis
    if face_app is None:
        print("Initializing FaceAnalysis...")
        providers = ["CUDAExecutionProvider", "CPUExecutionProvider"]
        face_app = FaceAnalysis(name="buffalo_l", root=str(CACHE_ROOT), providers=providers)
        face_app.prepare(ctx_id=0, det_size=(640, 640))
        print("FaceAnalysis initialized.")

    # Main Pipeline
    if pipe is None:
        print("Loading ControlNet...")
        controlnet = ControlNetModel.from_pretrained("InstantX/InstantID", subfolder="ControlNetModel", torch_dtype=dtype)
        
        print("Loading StableDiffusionPipeline...")
        pipe = StableDiffusionPipeline.from_single_file(BASE_CKPT, torch_dtype=dtype, safety_checker=None, use_safetensors=True, clip_skip=2)
        
        print("Moving pipeline to GPU...")
        pipe.to(device) # .to(device)をここで呼ぶ

        print("Loading VAE...")
        pipe.vae = AutoencoderKL.from_pretrained("stabilityai/sd-vae-ft-mse", torch_dtype=dtype).to(device)
        pipe.controlnet = controlnet
        
        print("Configuring Scheduler...")
        pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config, use_karras_sigmas=True, algorithm_type="sde-dpmsolver++")
        
        print("Loading IP-Adapter and LoRA...")
        pipe.load_ip_adapter("h94/IP-Adapter", subfolder="models", weight_name=IP_BIN_FILE.name)
        pipe.load_lora_weights(str(LORA_DIR), weight_name=LORA_FILE.name)
        
        pipe.set_ip_adapter_scale(0.65)
        print("Main pipeline initialized.")

    # Upscaler
    if upsampler is None and not UPSCALE_OK: # 一度失敗したら再試行しない
        print("Checking for Upscaler...")
        try:
            # cv2のインポートをここに追加
            import cv2
            from basicsr.archs.rrdb_arch import RRDBNet
            from realesrgan import RealESRGAN
            rrdb = RRDBNet(3, 3, 64, 23, 32, scale=8)
            upsampler = RealESRGAN(device, rrdb, scale=8)
            upsampler.load_weights(str(UPSCALE_DIR / "RealESRGAN_x8plus.pth"))
            UPSCALE_OK = True
            print("Upscaler initialized successfully.")
        except Exception as e:
            UPSCALE_OK = False # 失敗を記録
            print(f"Real-ESRGAN disabled → {e}")
    
    print("--- All pipelines ready ---")


# 4. Core generation logic
BASE_PROMPT = ("(masterpiece:1.2), best quality, ultra-realistic, RAW photo, 8k,\n""photo of {subject},\n""cinematic lighting, golden hour, rim light, shallow depth of field,\n""textured skin, high detail, shot on Canon EOS R5, 85 mm f/1.4, ISO 200,\n""<lora:ip-adapter-faceid-plusv2_sd15_lora:0.65>, (face),\n""(aesthetic:1.1), (cinematic:0.8)")
NEG_PROMPT = ("ng_deepnegative_v1_75t, CyberRealistic_Negative-neg, UnrealisticDream, ""(worst quality:2), (low quality:1.8), lowres, (jpeg artifacts:1.2), ""painting, sketch, illustration, drawing, cartoon, anime, cgi, render, 3d, ""monochrome, grayscale, text, logo, watermark, signature, username, ""(MajicNegative_V2:0.8), bad hands, extra digits, fused fingers, malformed limbs, ""missing arms, missing legs, (badhandv4:0.7), BadNegAnatomyV1-neg, skin blemishes, acnes, age spot, glans")

# 【変更点①】内部的な画像生成関数。@spaces.GPUデコレータを外す
def _generate_internal(face_img, subject, add_prompt, add_neg, cfg, ip_scale, steps, w, h, upscale, up_factor, progress=gr.Progress(track_tqdm=True)):
    # 初回呼び出し時にパイプラインを初期化
    initialize_pipelines()

    progress(0, desc="Generating image...")
    prompt = BASE_PROMPT.format(subject=(subject.strip() or "a beautiful 20yo woman"))
    if add_prompt: prompt += ", " + add_prompt
    neg = NEG_PROMPT + (", " + add_neg if add_neg else "")
    pipe.set_ip_adapter_scale(ip_scale)

    result = pipe(prompt=prompt, negative_prompt=neg, ip_adapter_image=face_img, image=face_img, controlnet_conditioning_scale=0.9, num_inference_steps=int(steps) + 5, guidance_scale=cfg, width=int(w), height=int(h)).images[0]
    
    if upscale and UPSCALE_OK:
        # cv2のインポートをここにも追加
        import cv2
        progress(0.8, desc="Upscaling...")
        up, _ = upsampler.enhance(cv2.cvtColor(np.array(result), cv2.COLOR_RGB2BGR), outscale=up_factor)
        result = Image.fromarray(cv2.cvtColor(up, cv2.COLOR_BGR2RGB))
    
    return result

# 【変更点②】@spaces.GPUデコレータを持つ新しいラッパー関数を定義
@spaces.GPU(duration=60)
def generate_gpu_wrapper(face_img, subject, add_prompt, add_neg, cfg, ip_scale, steps, w, h, upscale, up_factor, progress=gr.Progress(track_tqdm=True)):
    """
    Hugging Face SpacesプラットフォームにGPUを要求するためのラッパー関数。
    実際の処理は _generate_internal を呼び出して実行する。
    """
    return _generate_internal(face_img, subject, add_prompt, add_neg, cfg, ip_scale, steps, w, h, upscale, up_factor, progress)


# 【変更点③】GradioのUIから新しいラッパー関数を呼び出すように変更
def generate_ui(face_np, subject, add_prompt, add_neg, cfg, ip_scale, steps, w, h, upscale, up_factor, progress=gr.Progress(track_tqdm=True)):
    if face_np is None: raise gr.Error("顔画像をアップロードしてください。")
    # NumPy配列をPillow画像に変換
    face_img = Image.fromarray(face_np)
    # _generate_coreの代わりにgenerate_gpu_wrapperを呼び出す
    return generate_gpu_wrapper(face_img, subject, add_prompt, add_neg, cfg, ip_scale, steps, w, h, upscale, up_factor, progress)


# 5. Gradio UI Definition
with gr.Blocks() as demo:
    gr.Markdown("# InstantID – Beautiful Realistic Asians v7 (ZeroGPU)")
    with gr.Row():
        with gr.Column():
            face_in = gr.Image(label="顔写真",type="numpy")
            subj_in = gr.Textbox(label="被写体説明",placeholder="e.g. woman in black suit, smiling")
            add_in = gr.Textbox(label="追加プロンプト")
            addneg_in = gr.Textbox(label="追加ネガティブ")
            with gr.Accordion("詳細設定", open=False):
                ip_sld = gr.Slider(0,1.5,0.65,step=0.05,label="IP‑Adapter scale")
                cfg_sld = gr.Slider(1,15,6,step=0.5,label="CFG")
                step_sld = gr.Slider(10,50,20,step=1,label="Steps")
                w_sld = gr.Slider(512,1024,512,step=64,label="幅")
                h_sld = gr.Slider(512,1024,768,step=64,label="高さ")
                up_ck = gr.Checkbox(label="アップスケール",value=True)
                up_fac = gr.Slider(1,8,2,step=1,label="倍率")
            btn = gr.Button("生成",variant="primary")
        with gr.Column():
            out_img = gr.Image(label="結果")
            
    demo.queue()
    
    btn.click(
        fn=generate_ui, 
        inputs=[face_in,subj_in,add_in,addneg_in,cfg_sld,ip_sld,step_sld,w_sld,h_sld,up_ck,up_fac], 
        outputs=out_img
    )

# 6. FastAPI Mounting
app = FastAPI()

# 【変更点④】FastAPIのエンドポイントも新しいラッパー関数を呼び出すように変更
@app.post("/api/predict")
async def predict_endpoint(
    face_image: UploadFile = File(...),
    subject: str = Form("a woman"),
    add_prompt: str = Form(""),
    add_neg: str = Form(""),
    cfg: float = Form(6.0),
    ip_scale: float = Form(0.65),
    steps: int = Form(20),
    w: int = Form(512),
    h: int = Form(768),
    upscale: bool = Form(True),
    up_factor: float = Form(2.0)
):
    try:
        contents = await face_image.read()
        pil_image = Image.open(io.BytesIO(contents))
        
        # _generate_coreの代わりにgenerate_gpu_wrapperを呼び出す
        result_pil_image = generate_gpu_wrapper(
            pil_image, subject, add_prompt, add_neg, cfg, ip_scale, 
            steps, w, h, upscale, up_factor
        )

        buffered = io.BytesIO()
        result_pil_image.save(buffered, format="PNG")
        img_str = base64.b64encode(buffered.getvalue()).decode("utf-8")
        
        return {"image_base64": img_str}
    except Exception as e:
        traceback.print_exc()
        raise HTTPException(status_code=500, detail=str(e))

# GradioアプリをFastAPIアプリにマウント
app = gr.mount_gradio_app(app, demo, path="/")

print("Application startup script finished. Waiting for requests.")
if __name__ == "__main__":
    import os, time, socket, uvicorn

    def port_is_free(port: int) -> bool:
        with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:
            return s.connect_ex(("0.0.0.0", port)) != 0

    port = int(os.getenv("PORT", 7860))
    # ローカルでのテスト用にタイムアウトを短縮
     timeout_sec = 30
     poll_interval = 2
    
     t0 = time.time()
     while not port_is_free(port):
         waited = time.time() - t0
         if waited >= timeout_sec:
             raise RuntimeError(f"Port {port} is still busy after {timeout_sec}s")
         print(f"⚠️ Port {port} busy, retrying in {poll_interval}s …")
         time.sleep(poll_interval)
    
    # Hugging Face Spaces環境ではポートの競合は起こりにくいため、ポートチェックロジックを簡略化・無効化
    uvicorn.run(app, host="0.0.0.0", port=port, workers=1, log_level="info")