Update app.py

app.py

@@ -1,21 +1,20 @@
-# app.py — InstantID × Beautiful Realistic Asians v7 (ZeroGPU-ready, FastAPI + Gradio)
-# 2025-06-21 版
-#
-# ───────────────────────────────────────────────────────────────
-#  主な特徴
-#   • @spaces.GPU(duration=60) を公開名 generate_core() に付与
-#   • パイプラインは lazy-load で初回推論時に GPU へロード
-#   • モデル資産は /data または ~/.cache に永続化
-#   • Real-ESRGAN アップスケール (x4 / x8) オプション
-#   • Gradio UI + FastAPI REST を 1 プロセスで共存
-#   • Uvicorn 手動起動は不要（Spaces が自前で立てる）
-# ───────────────────────────────────────────────────────────────
-
-import os
-import io
-import base64
-import subprocess
-import traceback
+# app.py — InstantID × Beautiful Realistic Asians v7（ZeroGPU 対応フル版）
+# 2025-06-22
+
+##############################################################################
+# 0. 旧 API → 新 API 互換パッチ（必ず diffusers より前に実行）
+##############################################################################
+from huggingface_hub import hf_hub_download
+import huggingface_hub as _hf_hub
+
+# diffusers-0.27 が import する cached_download() を v0.28+ でも使えるように別名定義
+if not hasattr(_hf_hub, "cached_download"):
+    _hf_hub.cached_download = hf_hub_download
+
+##############################################################################
+# 1. 標準ライブラリ & 外部ライブラリ
+##############################################################################
+import os, io, base64, subprocess, traceback
 from pathlib import Path
 from typing import Optional
 
@@ -25,10 +24,6 @@ import gradio as gr
 import spaces
 from fastapi import FastAPI, UploadFile, File, Form, HTTPException
 from PIL import Image
-from huggingface_hub import hf_hub_download
-import huggingface_hub as _hf_hub
-if not hasattr(_hf_hub, "cached_download"):  # v0.26+ には存在しない
-    _hf_hub.cached_download = hf_hub_download
 
 from diffusers import (
     StableDiffusionControlNetPipeline,
@@ -41,10 +36,9 @@ from insightface.app import FaceAnalysis
 from basicsr.utils.download_util import load_file_from_url
 from realesrgan import RealESRGANer
 
-# ==============================================================
-# 0. キャッシュディレクトリとダウンローダ
-# ==============================================================
-
+##############################################################################
+# 2. キャッシュ／永続ディレクトリ
+##############################################################################
 PERSIST_BASE = Path("/data")
 CACHE_ROOT = (
     PERSIST_BASE / "instantid_cache"
@@ -57,9 +51,30 @@ UPSCALE_DIR = CACHE_ROOT / "realesrgan"
 for _p in (MODELS_DIR, LORA_DIR, UPSCALE_DIR):
     _p.mkdir(parents=True, exist_ok=True)
 
+##############################################################################
+# 3. モデル URL 一覧
+##############################################################################
+BRA_V7_URL = (
+    "https://huggingface.co/i0switch-assets/Beautiful_Realistic_Asians_v7/"
+    "resolve/main/beautiful_realistic_asians_v7_fp16.safetensors"
+)
+IP_ADAPTER_BIN_URL = (
+    "https://huggingface.co/h94/IP-Adapter/"
+    "resolve/main/ip-adapter-plus-face_sd15.bin"
+)
+IP_ADAPTER_LORA_URL = (
+    "https://huggingface.co/h94/IP-Adapter-FaceID/"
+    "resolve/main/ip-adapter-faceid-plusv2_sd15_lora.safetensors"
+)
+REALESRGAN_URL = (
+    "https://huggingface.co/aimagelab/realesrgan/"
+    "resolve/main/RealESRGAN_x4plus.pth"
+)
 
+##############################################################################
+# 4. ユーティリティ：堅牢ダウンロード
+##############################################################################
 def download(url: str, dst: Path, attempts: int = 2):
-    """単純リトライ付きダウンローダ（curl or basicsr fallback）"""
     if dst.exists():
         return dst
     for i in range(1, attempts + 1):
@@ -68,104 +83,73 @@ def download(url: str, dst: Path, attempts: int = 2):
             return dst
         except subprocess.CalledProcessError:
             print(f"[DL] Retry {i}/{attempts} failed: {url}")
-    # 最後に basicsr のダウンローダでフォールバック
     load_file_from_url(url=url, model_dir=str(dst.parent), file_name=dst.name)
     return dst
 
-
-# ==============================================================
-# 1. モデル URL 定義
-# ==============================================================
-
-BRA_V7_URL = (
-    "https://huggingface.co/i0switch-assets/Beautiful_Realistic_Asians_v7/"
-    "resolve/main/beautiful_realistic_asians_v7_fp16.safetensors"
-)
-IP_ADAPTER_BIN_URL = (
-    "https://huggingface.co/h94/IP-Adapter/resolve/main/ip-adapter-plus-face_sd15.bin"
-)
-IP_ADAPTER_LORA_URL = (
-    "https://huggingface.co/h94/IP-Adapter-FaceID/resolve/main/"
-    "ip-adapter-faceid-plusv2_sd15_lora.safetensors"
-)
-REALESRGAN_URL = (
-    "https://huggingface.co/aimagelab/realesrgan/resolve/main/RealESRGAN_x4plus.pth"
-)
-
-# ==============================================================
-# 2. グローバル変数（lazy-load される）
-# ==============================================================
-
+##############################################################################
+# 5. グローバル（lazy-load される）
+##############################################################################
 pipe: Optional[StableDiffusionControlNetPipeline] = None
 face_analyser: Optional[FaceAnalysis] = None
 upsampler: Optional[RealESRGANer] = None
 
-# ==============================================================
-# 3. パイプライン初期化
-# ==============================================================
-
-
+##############################################################################
+# 6. パイプライン初期化
+##############################################################################
 def initialize_pipelines():
+    """最初の推論時に一度だけ呼び出す"""
     global pipe, face_analyser, upsampler
-
     if pipe is not None:
-        return  # 既に初期化済み
+        return
 
     print("[INIT] Downloading model assets …")
 
-    # ---- 3-1. 基本モデル & IP-Adapter ----
+    # (1) ベースモデル & IP-Adapter
     bra_ckpt = download(BRA_V7_URL, MODELS_DIR / "bra_v7.safetensors")
     ip_bin = download(IP_ADAPTER_BIN_URL, MODELS_DIR / "ip_adapter.bin")
     ip_lora = download(IP_ADAPTER_LORA_URL, LORA_DIR / "ip_adapter_faceid.lora")
 
-    # ---- 3-2. ControlNet (InstantID) ----
+    # (2) ControlNet (InstantID)
     controlnet = ControlNetModel.from_pretrained(
         "InstantID/ControlNet-Mediapipe-Face",
         torch_dtype=torch.float16,
         cache_dir=str(MODELS_DIR),
     )
 
-    # ---- 3-3. Diffusers パイプライン ----
-    pipe_local_files_only = {
-        "controlnet": controlnet,
-        "vae": AutoencoderKL.from_pretrained(
+    # (3) Diffusers Pipeline
+    pipe_tmp = StableDiffusionControlNetPipeline.from_pretrained(
+        "runwayml/stable-diffusion-v1-5",
+        controlnet=controlnet,
+        vae=AutoencoderKL.from_pretrained(
             "stabilityai/sd-vae-ft-mse", torch_dtype=torch.float16
         ),
-        "torch_dtype": torch.float16,
-        "safety_checker": None,
-    }
-    pipe_base = "runwayml/stable-diffusion-v1-5"
-    pipe_kwargs = dict(
-        local_files_only=False,
+        torch_dtype=torch.float16,
         cache_dir=str(MODELS_DIR),
-        load_safety_checker=False,
-    )
-    pipe_tmp = StableDiffusionControlNetPipeline.from_pretrained(
-        pipe_base, **pipe_local_files_only, **pipe_kwargs
+        safety_checker=None,
     )
     pipe_tmp.scheduler = DPMSolverMultistepScheduler.from_pretrained(
-        pipe_base, subfolder="scheduler", cache_dir=str(MODELS_DIR)
-    )
-    # LoRA / IP-Adapter
-    pipe_tmp.load_ip_adapter(
-        ip_bin,
-        subfolder=None,
-        weight_name=None,
+        "runwayml/stable-diffusion-v1-5",
+        subfolder="scheduler",
+        cache_dir=str(MODELS_DIR),
     )
+    # IP-Adapter：LoRA を適用
+    pipe_tmp.load_ip_adapter(ip_bin)
     ip_layers = AttnProcsLayers(pipe_tmp.unet.attn_processors)
-    ip_layers.load_lora_weights(ip_lora, adapter_name="ip_faceid", safe_load=True)
+    ip_layers.load_lora_weights(
+        ip_lora, adapter_name="ip_faceid", safe_load=True
+    )
     pipe_tmp.set_adapters(["ip_faceid"], adapter_weights=[0.6])
     pipe_tmp.to("cuda")
 
     pipe = pipe_tmp
 
-    # ---- 3-4. InsightFace ----
+    # (4) InsightFace
     face_analyser = FaceAnalysis(
         name="buffalo_l", root=str(MODELS_DIR), providers=["CUDAExecutionProvider"]
     )
     face_analyser.prepare(ctx_id=0, det_size=(640, 640))
 
-    # ---- 3-5. Real-ESRGAN ----
+    # (5) Real-ESRGAN
     esrgan_ckpt = download(REALESRGAN_URL, UPSCALE_DIR / "realesrgan_x4plus.pth")
     upsampler = RealESRGANer(
         scale=4,
@@ -179,26 +163,22 @@ def initialize_pipelines():
 
     print("[INIT] Pipelines ready.")
 
-
-# ==============================================================
-# 4. プロンプト設定
-# ==============================================================
-
+##############################################################################
+# 7. プロンプトテンプレ
+##############################################################################
 BASE_PROMPT = (
-    "(masterpiece:1.2), best quality, ultra-realistic, 8k, RAW photo, "
+    "(masterpiece:1.2), best quality, ultra-realistic, RAW photo, 8k, "
     "cinematic lighting, textured skin, "
 )
 NEG_PROMPT = (
     "verybadimagenegative_v1.3, ng_deepnegative_v1_75t, "
     "(worst quality:2), (low quality:2), lowres, blurry, bad anatomy, "
-    "bad hands, extra digits, cropped, watermark, signature"
+    "bad hands, extra digits, watermark, signature"
 )
 
-# ==============================================================
-# 5. 生成コア関数（GPU を掴む）
-# ==============================================================
-
-
+##############################################################################
+# 8. 生成関数（GPU を掴む）
+##############################################################################
 @spaces.GPU(duration=60)
 def generate_core(
     face_img: Image.Image,
@@ -214,17 +194,14 @@ def generate_core(
     up_factor: int = 4,
     progress: gr.Progress = gr.Progress(track_tqdm=True),
 ):
-    """
-    メイン生成関数
-    """
     try:
         if pipe is None:
             initialize_pipelines()
 
-        face_np = np.array(face_img)
-        face_info = face_analyser.get(face_np)
-        if len(face_info) == 0:
-            raise ValueError("顔が検出できませんでした。")
+        np_face = np.array(face_img)
+        faces = face_analyser.get(np_face)
+        if len(faces) == 0:
+            raise ValueError("顔が検出できませんでした。別の画像でお試しください。")
 
         pipe.set_adapters(["ip_faceid"], adapter_weights=[ip_scale])
 
@@ -243,8 +220,7 @@ def generate_core(
         ).images[0]
 
         if upscale and upsampler is not None:
-            scale = 4 if up_factor == 4 else 8
-            upsampler.scale = scale
+            upsampler.scale = 4 if up_factor == 4 else 8
             result, _ = upsampler.enhance(np.array(result))
             result = Image.fromarray(result)
 
@@ -254,17 +230,15 @@ def generate_core(
         traceback.print_exc()
         raise e
 
-
-# ==============================================================
-# 6. Gradio UI
-# ==============================================================
-
+##############################################################################
+# 9. Gradio UI
+##############################################################################
 with gr.Blocks(title="InstantID × BRA v7 (ZeroGPU)") as demo:
     gr.Markdown("## InstantID × Beautiful Realistic Asians v7")
     with gr.Row():
         face_img = gr.Image(type="pil", label="Face ID", sources=["upload"])
         subject = gr.Textbox(
-            label="被写体説明（例: '30代日本人女性、黒髪セミロング'）", interactive=True
+            label="被写体説明（例: 30代日本人女性、黒髪セミロング）", interactive=True
         )
     add_prompt = gr.Textbox(label="追加プロンプト", interactive=True)
     add_neg = gr.Textbox(label="追加ネガティブ", interactive=True)
@@ -279,7 +253,6 @@ with gr.Blocks(title="InstantID × BRA v7 (ZeroGPU)") as demo:
         upscale = gr.Checkbox(label="Real-ESRGAN Upscale", value=False)
         up_factor = gr.Radio([4, 8], value=4, label="Upscale Factor")
     run_btn = gr.Button("Generate")
-
     output_img = gr.Image(type="pil", label="Result")
 
     run_btn.click(
@@ -301,13 +274,11 @@ with gr.Blocks(title="InstantID × BRA v7 (ZeroGPU)") as demo:
         show_progress=True,
     )
 
-# ==============================================================
-# 7. FastAPI エンドポイント
-# ==============================================================
-
+##############################################################################
+# 10. FastAPI エンドポイント
+##############################################################################
 app = FastAPI()
 
-
 @app.post("/api/generate")
 async def api_generate(
     subject: str = Form(...),
@@ -342,10 +313,7 @@ async def api_generate(
         traceback.print_exc()
         raise HTTPException(status_code=500, detail=str(e))
 
-
-# ==============================================================
-# 8. Launch
-# ==============================================================
-
-# Spaces が自動で Uvicorn を起動するため、手動起動は不要。
-demo.queue(concurrency_count=2).launch(share=False)
+##############################################################################
+# 11. Launch（Gradio が自動で Uvicorn を立ち上げる）
+##############################################################################
+demo.queue(default_concurrency_limit=2).launch(share=False)