Update app.py

app.py

@@ -1,5 +1,5 @@
-# app.py — InstantID × Beautiful Realistic Asians v7（ZeroGPU 対応・CrucibleAI ControlNet）
-# 2025-06-22
+# app.py — InstantID × Beautiful Realistic Asians v7（ZeroGPU / ControlNetMediaPipeFace）
+# 2025-06-22 版
 
 ##############################################################################
 # 0. 旧 API → 新 API 互換パッチ（必ず diffusers import の前に置く）
@@ -7,9 +7,9 @@
 from huggingface_hub import hf_hub_download
 import huggingface_hub as _hf_hub
 
-# diffusers-0.27 は cached_download() を呼び出すため、v0.28+ でも使えるよう注入
+# diffusers-0.27 は cached_download() を呼び出すため、HF-Hub ≥0.28 でも使えるように注入
 if not hasattr(_hf_hub, "cached_download"):
-    _hf_hub.cached_download = hf_hub_download
+    _hf_hub.cached_download = hf_hub_download  # :contentReference[oaicite:1]{index=1}
 
 ##############################################################################
 # 1. 標準 & 外部ライブラリ
@@ -37,7 +37,7 @@ from basicsr.utils.download_util import load_file_from_url
 from realesrgan import RealESRGANer
 
 ##############################################################################
-# 2. キャッシュ & パス
+# 2. キャッシュ & 永続パス
 ##############################################################################
 PERSIST_BASE = Path("/data")
 CACHE_ROOT = (
@@ -46,43 +46,54 @@ CACHE_ROOT = (
     else Path.home() / ".cache" / "instantid_cache"
 )
 MODELS_DIR = CACHE_ROOT / "models"
-LORA_DIR = CACHE_ROOT / "lora"
+LORA_DIR   = CACHE_ROOT / "lora"
 UPSCALE_DIR = CACHE_ROOT / "realesrgan"
-for _p in (MODELS_DIR, LORA_DIR, UPSCALE_DIR):
-    _p.mkdir(parents=True, exist_ok=True)
+for p in (MODELS_DIR, LORA_DIR, UPSCALE_DIR):
+    p.mkdir(parents=True, exist_ok=True)
 
 ##############################################################################
-# 3. モデル URL 一覧
+# 3. モデル識別子 & ファイル名
 ##############################################################################
-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"
-)
+# すべて HF Hub 側にバイナリがあるため、curl ではなく hf_hub_download() を推奨
+BRA_REPO    = "i0switch-assets/Beautiful_Realistic_Asians_v7"
+BRA_FILE    = "beautiful_realistic_asians_v7_fp16.safetensors"
+
+IP_REPO     = "h94/IP-Adapter"
+IP_FILE_BIN = "ip-adapter-plus-face_sd15.bin"       # Git LFS バイナリ :contentReference[oaicite:2]{index=2}
+
+IP_LORA_REPO = "h94/IP-Adapter-FaceID"
+IP_FILE_LORA = "ip-adapter-faceid-plusv2_sd15_lora.safetensors"  # Git LFS バイナリ
+
+CN_REPO   = "CrucibleAI/ControlNetMediaPipeFace"  # 公開・無認証で DL 可 :contentReference[oaicite:3]{index=3}
+CN_FOLDER = "diffusion_sd15"                      # SD-1.5 用フォルダ :contentReference[oaicite:4]{index=4}
+
+REALESRGAN_REPO = "aimagelab/realesrgan"
+REALESRGAN_FILE = "RealESRGAN_x4plus.pth"
 
 ##############################################################################
-# 4. ダウンローダ
+# 4. ダウンローダ（HF Hub 優先）
 ##############################################################################
-def download(url: str, dst: Path, attempts: int = 2):
+def dl_hf(repo: str, filename: str, subfolder: Optional[str] = None) -> Path:
+    """HF Hub から大容量バイナリを安全に取得（Git LFS ポインタ問題を回避）"""
+    return Path(
+        hf_hub_download(
+            repo_id=repo,
+            filename=filename,
+            subfolder=subfolder,
+            cache_dir=str(MODELS_DIR),
+        )
+    )
+
+def dl_http(url: str, dst: Path):
+    """小さなファイルのみ curl で取得（retry 付）"""
     if dst.exists():
         return dst
-    for i in range(1, attempts + 1):
+    for _ in range(2):
         try:
             subprocess.check_call(["curl", "-L", "-o", str(dst), url])
             return dst
         except subprocess.CalledProcessError:
-            print(f"[DL] Retry {i}/{attempts} failed: {url}")
+            pass
     load_file_from_url(url=url, model_dir=str(dst.parent), file_name=dst.name)
     return dst
 
@@ -103,23 +114,18 @@ def initialize_pipelines():
 
     print("[INIT] Downloading model assets …")
 
-    # 6-1 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")
-
-    # 6-2 ControlNet（CrucibleAI / diffusion_sd15）
-    controlnet = ControlNetModel.from_pretrained(
-        "CrucibleAI/ControlNetMediaPipeFace",  # 公開リポジトリ :contentReference[oaicite:0]{index=0}
-        subfolder="diffusion_sd15",            # SD-1.5 用フォルダ :contentReference[oaicite:1]{index=1}
-        torch_dtype=torch.float16,
-        cache_dir=str(MODELS_DIR),
+    # 6-1 主要モデル
+    bra_ckpt = dl_hf(BRA_REPO, BRA_FILE)
+    ip_bin   = dl_hf(IP_REPO,  IP_FILE_BIN)
+    ip_lora  = dl_hf(IP_LORA_REPO, IP_FILE_LORA)
+    cn_model = ControlNetModel.from_pretrained(
+        CN_REPO, subfolder=CN_FOLDER, torch_dtype=torch.float16, cache_dir=str(MODELS_DIR)
     )
 
-    # 6-3 Diffusers パイプライン
+    # 6-2 Diffusers パイプライン
     pipe_tmp = StableDiffusionControlNetPipeline.from_pretrained(
         "runwayml/stable-diffusion-v1-5",
-        controlnet=controlnet,
+        controlnet=cn_model,
         vae=AutoencoderKL.from_pretrained(
             "stabilityai/sd-vae-ft-mse", torch_dtype=torch.float16
         ),
@@ -133,16 +139,12 @@ def initialize_pipelines():
         cache_dir=str(MODELS_DIR),
     )
 
-    # 6-4 **IP-Adapter 読み込み（API 仕様は positional 三つ）**
-    #     diffusers-0.27.2 では subfolder / weight_name が必須 :contentReference[oaicite:2]{index=2}
-    ip_dir = ip_bin.parent
+    # 6-3 IP-Adapter ロード（必須 3 引数） :contentReference[oaicite:5]{index=5}
     pipe_tmp.load_ip_adapter(
-        str(ip_dir),          # path or repo id
-        "",                   # subfolder（ファイル直下なので空文字列）
+        str(ip_bin.parent),   # repo_or_path
+        "",                   # subfolder（直下なので空文字）
         ip_bin.name           # weight_name
     )
-
-    # IP-Adapter の追加 LoRA を合流
     AttnProcsLayers(pipe_tmp.unet.attn_processors).load_lora_weights(
         ip_lora, adapter_name="ip_faceid", safe_load=True
     )
@@ -151,22 +153,19 @@ def initialize_pipelines():
 
     pipe = pipe_tmp
 
-    # 6-5 InsightFace
+    # 6-4 InsightFace
     face_analyser = FaceAnalysis(
         name="buffalo_l", root=str(MODELS_DIR), providers=["CUDAExecutionProvider"]
     )
     face_analyser.prepare(ctx_id=0, det_size=(640, 640))
 
-    # 6-6 Real-ESRGAN
-    esrgan_ckpt = download(REALESRGAN_URL, UPSCALE_DIR / "realesrgan_x4plus.pth")
+    # 6-5 Real-ESRGAN
+    re_ckpt = dl_hf(REALESRGAN_REPO, REALESRGAN_FILE)
     upsampler = RealESRGANer(
         scale=4,
-        model_path=str(esrgan_ckpt),
+        model_path=str(re_ckpt),
         half=True,
-        tile=512,
-        tile_pad=10,
-        pre_pad=0,
-        gpu_id=0,
+        tile=512, tile_pad=10, pre_pad=0, gpu_id=0
     )
 
     print("[INIT] Pipelines ready.")
@@ -185,9 +184,9 @@ NEG_PROMPT = (
 )
 
 ##############################################################################
-# 8. 生成関数（GPU アタッチ）
+# 8. 生成コア（GPU アタッチ）
 ##############################################################################
-@spaces.GPU(duration=60)
+@spaces.GPU(duration=60)  # ZeroGPU で 60 s まで実行可 :contentReference[oaicite:6]{index=6}
 def generate_core(
     face_img: Image.Image,
     subject: str,
@@ -206,14 +205,13 @@ def generate_core(
         if pipe is None:
             initialize_pipelines()
 
-        faces = face_analyser.get(np.array(face_img))
-        if len(faces) == 0:
-            raise ValueError("顔が検出できませんでした。別の画像をお試しください。")
+        if len(face_analyser.get(np.array(face_img))) == 0:
+            raise ValueError("顔が検出できません。別の画像でお試しください。")
 
         pipe.set_adapters(["ip_faceid"], adapter_weights=[ip_scale])
 
-        prompt = BASE_PROMPT + subject + ", " + add_prompt
-        negative = NEG_PROMPT + ", " + add_neg
+        prompt   = BASE_PROMPT + subject + ", " + add_prompt
+        negative = NEG_PROMPT  + ", " + add_neg
 
         result = pipe(
             prompt=prompt,
@@ -222,11 +220,10 @@ def generate_core(
             guidance_scale=float(cfg),
             image=face_img,
             control_image=None,
-            width=int(w),
-            height=int(h),
+            width=int(w), height=int(h),
         ).images[0]
 
-        if upscale and upsampler is not None:
+        if upscale:
             upsampler.scale = 4 if up_factor == 4 else 8
             result, _ = upsampler.enhance(np.array(result))
             result = Image.fromarray(result)
@@ -244,41 +241,27 @@ 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
-        )
+        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)
     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)
+    add_neg    = gr.Textbox(label="追加ネガティブ", interactive=True)
     with gr.Row():
-        cfg = gr.Slider(1, 20, value=7.5, step=0.5, label="CFG Scale")
+        cfg      = gr.Slider(1, 20, value=7.5, step=0.5, label="CFG Scale")
         ip_scale = gr.Slider(0.1, 1.0, value=0.6, step=0.05, label="IP-Adapter Weight")
     with gr.Row():
         steps = gr.Slider(10, 50, value=30, step=1, label="Steps")
-        w = gr.Slider(512, 1024, value=768, step=64, label="Width")
-        h = gr.Slider(512, 1024, value=768, step=64, label="Height")
+        w     = gr.Slider(512, 1024, value=768, step=64, label="Width")
+        h     = gr.Slider(512, 1024, value=768, step=64, label="Height")
     with gr.Row():
-        upscale = gr.Checkbox(label="Real-ESRGAN Upscale", value=False)
+        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   = gr.Button("Generate")
+    output_im = gr.Image(type="pil", label="Result")
 
     run_btn.click(
         fn=generate_core,
-        inputs=[
-            face_img,
-            subject,
-            add_prompt,
-            add_neg,
-            cfg,
-            ip_scale,
-            steps,
-            w,
-            h,
-            upscale,
-            up_factor,
-        ],
-        outputs=output_img,
-        show_progress=True,
+        inputs=[face_img, subject, add_prompt, add_neg,
+                cfg, ip_scale, steps, w, h, upscale, up_factor],
+        outputs=output_im, show_progress=True
     )
 
 ##############################################################################
@@ -297,25 +280,10 @@ async def api_generate(
     file: UploadFile = File(...),
 ):
     try:
-        img_bytes = await file.read()
-        pil = Image.open(io.BytesIO(img_bytes)).convert("RGB")
-        res = generate_core(
-            face_img=pil,
-            subject=subject,
-            add_prompt="",
-            add_neg="",
-            cfg=cfg,
-            ip_scale=ip_scale,
-            steps=steps,
-            w=w,
-            h=h,
-            upscale=False,
-            up_factor=4,
-        )
-        buf = io.BytesIO()
-        res.save(buf, format="PNG")
-        b64 = base64.b64encode(buf.getvalue()).decode()
-        return {"image": f"data:image/png;base64,{b64}"}
+        img = Image.open(io.BytesIO(await file.read())).convert("RGB")  # noqa
+        res = generate_core(img, subject, "", "", cfg, ip_scale, steps, w, h, False, 4)
+        buf = io.BytesIO(); res.save(buf, format="PNG")
+        return {"image": "data:image/png;base64," + base64.b64encode(buf.getvalue()).decode()}
     except Exception as e:
         traceback.print_exc()
         raise HTTPException(status_code=500, detail=str(e))
@@ -323,4 +291,4 @@ async def api_generate(
 ##############################################################################
 # 11. Launch（Gradio が自動で Uvicorn を起動）
 ##############################################################################
-demo.queue(default_concurrency_limit=2).launch(share=False)
+demo.queue(default_concurrency_limit=2).launch(share=False)  # :contentReference[oaicite:7]{index=7}