Update app.py

app.py

@@ -1,15 +1,10 @@
 # app.py — InstantID SDXL (officiel) + IP-Adapter Style (optionnel, rendu 2D)
-# Hugging Face Space ready
 
-# 0) Environnement AVANT imports
 import os, sys
 os.environ["OMP_NUM_THREADS"] = "4"
 os.environ.setdefault("HF_HUB_ENABLE_HF_TRANSFER", "1")
-
-# rendre importable ./instantid (pipeline officielle à placer ici)
 sys.path.insert(0, os.path.abspath("./instantid"))
 
-# 1) Imports
 import traceback, importlib.util
 import torch, gradio as gr
 from PIL import Image, ImageOps, ImageDraw
@@ -20,7 +15,6 @@ from insightface.app import FaceAnalysis
 DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
 DTYPE  = torch.float16 if DEVICE == "cuda" else torch.float32
 
-# 2) Chemins & Hub (poids InstantID officiels + IP-Adapter Style SDXL)
 ASSETS_REPO = "InstantX/InstantID"
 CHECKPOINTS_DIR  = "./checkpoints"
 CN_LOCAL_DIR     = os.path.join(CHECKPOINTS_DIR, "ControlNetModel")
@@ -30,15 +24,12 @@ IP_STYLE_REPO      = "h94/IP-Adapter"
 IP_STYLE_SUBFOLDER = "sdxl_models"
 IP_STYLE_WEIGHT    = "ip-adapter_sdxl.bin"
 
-# Modèle de base (remplaçable par un checkpoint anime)
 BASE_MODEL = "stabilityai/stable-diffusion-xl-base-1.0"
 
-# 3) Téléchargements sûrs
 def safe_download(repo, filename, local_dir, min_bytes, label, subfolder=None):
     os.makedirs(local_dir, exist_ok=True)
     local_path = os.path.join(local_dir, os.path.basename(filename))
     if os.path.exists(local_path) and os.path.getsize(local_path) < min_bytes:
-        print(f"⚠️ {label} corrompu ({os.path.getsize(local_path)} bytes) → suppression")
         try: os.remove(local_path)
         except Exception: pass
     path = hf_hub_download(
@@ -51,22 +42,19 @@ def safe_download(repo, filename, local_dir, min_bytes, label, subfolder=None):
         subfolder=subfolder,
     )
     size = os.path.getsize(path)
-    print(f"✅ {label} téléchargé ({size/1e6:.1f} MB)")
     if size < min_bytes:
         raise RuntimeError(f"Téléchargement incomplet de {label} (taille: {size} bytes).")
+    print(f"✅ {label} téléchargé ({size/1e6:.1f} MB)")
     return path
 
 def ensure_assets_or_download():
     os.makedirs(CHECKPOINTS_DIR, exist_ok=True)
     os.makedirs(CN_LOCAL_DIR, exist_ok=True)
-    # IdentityNet (ControlNet) + ip-adapter (InstantID)
     safe_download(ASSETS_REPO, "ControlNetModel/config.json", CHECKPOINTS_DIR, 1_000, "IdentityNet config")
     safe_download(ASSETS_REPO, "ControlNetModel/diffusion_pytorch_model.safetensors", CHECKPOINTS_DIR, 100_000_000, "IdentityNet weights")
     safe_download(ASSETS_REPO, "ip-adapter.bin", CHECKPOINTS_DIR, 100_000_000, "IP-Adapter (InstantID)")
-    # IP-Adapter Style (SDXL) — optionnel
     safe_download(IP_STYLE_REPO, IP_STYLE_WEIGHT, CHECKPOINTS_DIR, 20_000_000, "IP-Adapter Style (SDXL)", subfolder=IP_STYLE_SUBFOLDER)
 
-# 4) Import dynamique de la pipeline SDXL officielle
 def import_pipeline_or_fail():
     candidates = [
         "./instantid/pipeline_stable_diffusion_xl_instantid_full.py",
@@ -77,7 +65,6 @@ def import_pipeline_or_fail():
         raise RuntimeError("❌ Pipeline manquante. Place `pipeline_stable_diffusion_xl_instantid_full.py` dans ./instantid/")
     if os.path.getsize(pipeline_file) < 1024:
         raise RuntimeError("❌ Pipeline trop petite (vide ?). Utilise la version SDXL officielle.")
-
     spec = importlib.util.spec_from_file_location("instantid_pipeline", pipeline_file)
     mod  = importlib.util.module_from_spec(spec)
     spec.loader.exec_module(mod)
@@ -88,7 +75,6 @@ def import_pipeline_or_fail():
     avail = [n for n, o in vars(mod).items() if isinstance(o, type)]
     raise RuntimeError("❌ Aucune classe pipeline InstantID trouvée. Classes dispo: " + ", ".join(avail))
 
-# 5) util — dessin landmarks (kps)
 def draw_kps_local(img_pil, kps):
     w, h = img_pil.size
     out = Image.new("RGB", (w, h), "white")
@@ -98,30 +84,23 @@ def draw_kps_local(img_pil, kps):
         d.ellipse((x - r, y - r, x + r, y + r), fill="black")
     return out
 
-# 6) Chargement pipeline
 load_logs = []
 HAS_STYLE_ADAPTER = False
-
 try:
-    # a) pipeline
     SDXLInstantID = import_pipeline_or_fail()
     ensure_assets_or_download()
 
-    # b) IdentityNet (ControlNet)
     controlnet_identitynet = ControlNetModel.from_pretrained(CN_LOCAL_DIR, torch_dtype=DTYPE)
-
     pipe = SDXLInstantID.from_pretrained(
         BASE_MODEL,
-        controlnet=controlnet_identitynet,              # objet unique
+        controlnet=controlnet_identitynet,
         torch_dtype=DTYPE,
         safety_checker=None,
         feature_extractor=None,
     ).to(DEVICE)
 
-    # c) IP-Adapter InstantID (identité)
     pipe.load_ip_adapter_instantid(IP_ADAPTER_LOCAL)
 
-    # d) IP-Adapter Style SDXL (optionnel), nommé "style"
     try:
         pipe.load_ip_adapter(
             IP_STYLE_REPO,
@@ -129,187 +108,4 @@ try:
             weight_name=IP_STYLE_WEIGHT,
             adapter_name="style",
         )
-        load_logs.append("✅ IP-Adapter Style (SDXL) chargé (adapter_name='style').")
-        HAS_STYLE_ADAPTER = True
-    except Exception as e:
-        load_logs.append(f"ℹ️ IP-Adapter Style non chargé: {e}")
-
-    # e) devices
-    if DEVICE == "cuda":
-        if hasattr(pipe, "image_proj_model"): pipe.image_proj_model.to("cuda")
-        if hasattr(pipe, "unet"):             pipe.unet.to("cuda")
-
-    load_logs.append("✅ InstantID prêt.")
-except Exception:
-    load_logs += ["❌ ERREUR au chargement:", traceback.format_exc()]
-    pipe = None
-
-if pipe is None:
-    raise RuntimeError("Échec de chargement du pipeline.\n" + "\n".join(load_logs))
-
-# 7) InsightFace (antelopev2 → buffalo_l)
-def load_face_analyser():
-    errors = []
-    for name in ("antelopev2", "buffalo_l"):
-        try:
-            fa = FaceAnalysis(name=name, root="./models", providers=["CPUExecutionProvider"])
-            fa.prepare(ctx_id=0, det_size=(640, 640))
-            print(f"✅ InsightFace chargé: {name}")
-            return fa
-        except Exception as e:
-            errors.append(f"{name}: {e}")
-            print(f"⚠️ InsightFace échec {name} → {e}")
-    raise RuntimeError("Echec chargement InsightFace. Détails: " + " | ".join(errors))
-
-fa = load_face_analyser()
-
-def extract_face_embed_and_kps(pil_img):
-    """
-    Retourne:
-      - face_emb (torch.Tensor de forme [1, D] sur DEVICE/DTYPE) — requis pour UNet (encoder_hid_dim_type='ip_image_proj')
-      - kps_img  (PIL.Image) — landmarks pour IdentityNet
-    """
-    import numpy as np, cv2
-    img_cv2 = cv2.cvtColor(np.array(pil_img.convert("RGB")), cv2.COLOR_RGB2BGR)
-    faces = fa.get(img_cv2)
-    if not faces:
-        raise ValueError("Aucun visage détecté dans la photo.")
-    face = faces[-1]
-
-    # (1) Embedding InsightFace -> torch [1, D] sur bon device/dtype
-    emb_np = face["embedding"]
-    if not isinstance(emb_np, np.ndarray):
-        emb_np = np.asarray(emb_np, dtype="float32")
-    if emb_np.ndim == 1:
-        emb_np = emb_np[None, ...]  # (1, D)
-    face_emb = torch.from_numpy(emb_np).to(device=DEVICE, dtype=DTYPE)
-
-    # (2) Landmarks -> image kps
-    kps_img = draw_kps_local(pil_img, face["kps"])
-    return face_emb, kps_img
-
-# 8) Génération
-def generate(face_image, style_image, prompt, negative_prompt,
-             identity_strength, adapter_strength, style_strength,
-             steps, cfg, width, height, seed):
-    try:
-        if face_image is None:
-            return None, "Merci d'ajouter une photo visage.", "\n".join(load_logs)
-
-        gen = None if seed is None or int(seed) < 0 else torch.Generator(device=DEVICE).manual_seed(int(seed))
-
-        # visage → carré 512 pour détection stable
-        face = ImageOps.exif_transpose(face_image).convert("RGB")
-        ms = min(face.size); x = (face.width - ms) // 2; y = (face.height - ms) // 2
-        face_sq = face.crop((x, y, x + ms, y + ms)).resize((512, 512), Image.Resampling.LANCZOS)
-
-        # InsightFace : embedding (torch [1,D]) + landmarks
-        face_emb, kps_img = extract_face_embed_and_kps(face_sq)
-
-        # IP-Adapter scales
-        try:
-            if HAS_STYLE_ADAPTER and style_image is not None:
-                pipe.set_ip_adapter_scale({"instantid": float(adapter_strength), "style": float(style_strength)})
-            else:
-                pipe.set_ip_adapter_scale(float(adapter_strength))
-        except Exception as e:
-            print(f"ℹ️ set_ip_adapter_scale ignoré: {e}")
-
-        # compat multi-ControlNet (même si on en a qu’un)
-        cn = getattr(pipe, "controlnet", None)
-        if isinstance(cn, (list, tuple)):
-            n_cn = len(cn)
-        else:
-            try: n_cn = len(cn)
-            except Exception: n_cn = 1
-
-        image_arg = [kps_img] * n_cn if n_cn > 1 else ([kps_img] if isinstance(cn, (list, tuple)) else kps_img)
-        scale_val = float(identity_strength)
-        scale_arg = [scale_val] * n_cn if n_cn > 1 else ([scale_val] if isinstance(cn, (list, tuple)) else scale_val)
-
-        # kwargs d’inférence — IMPORTANT: image_embeds est un torch Tensor [1,D]
-        gen_kwargs = dict(
-            prompt=(prompt or "").strip(),
-            negative_prompt=(negative_prompt or "").strip(),
-            image=image_arg,                         # IdentityNet (landmarks)
-            image_embeds=face_emb,                   # ← torch.Tensor [1, D] sur DEVICE/DTYPE
-            controlnet_conditioning_scale=scale_arg,
-            num_inference_steps=int(steps),
-            guidance_scale=float(cfg),
-            width=int(width),
-            height=int(height),
-            generator=gen,
-        )
-
-        # passer l’image de style à l’IP-Adapter Style (si dispo + fournie)
-        if HAS_STYLE_ADAPTER and style_image is not None:
-            try:
-                gen_kwargs["ip_adapter_image"] = ImageOps.exif_transpose(style_image).convert("RGB")
-            except Exception as e:
-                print(f"ℹ️ ip_adapter_image ignoré: {e}")
-
-        images = pipe(**gen_kwargs).images
-        return images[0], "", "\n".join(load_logs)
-
-    except torch.cuda.OutOfMemoryError:
-        return None, "CUDA OOM: baisse la résolution ou les steps.", "\n".join(load_logs)
-    except Exception:
-        return None, "Erreur:\n" + traceback.format_exc(), "\n".join(load_logs)
-
-# 9) UI
-EX_PROMPT = (
-    "one piece style, Eiichiro Oda style, anime portrait, upper body, pirate outfit, "
-    "clean lineart, cel shading, vibrant colors, expressive eyes, dynamic composition, simple background"
-)
-EX_NEG = (
-    "realistic, photo, photorealistic, skin pores, complex lighting, "
-    "low quality, worst quality, lowres, blurry, noisy, watermark, text, logo, jpeg artifacts, "
-    "bad anatomy, deformed, multiple faces, nsfw"
-)
-
-with gr.Blocks(css="footer{display:none !important}") as demo:
-    gr.Markdown("# 🏴‍☠️ InstantID SDXL + IP-Adapter Style (2D) — visage → perso One Piece")
-
-    with gr.Row():
-        with gr.Column():
-            face_image  = gr.Image(type="pil", label="Photo visage (obligatoire)", height=260)
-            style_image = gr.Image(type="pil", label="Image de style (optionnel)", height=260)
-
-            gr.Markdown("Astuce : charge un poster/planche One Piece (ou visuel manga) pour forcer le rendu 2D via IP-Adapter Style.")
-
-            prompt   = gr.Textbox(label="Prompt", value=EX_PROMPT, lines=3)
-            negative = gr.Textbox(label="Negative Prompt", value=EX_NEG, lines=3)
-
-            with gr.Row():
-                identity_strength = gr.Slider(0.2, 1.5, 0.95, 0.05, label="Fidélité visage (IdentityNet)")
-                adapter_strength  = gr.Slider(0.1, 1.5, 0.85, 0.05, label="Détails anime (InstantID)")
-
-            style_strength = gr.Slider(0.1, 1.5, 0.95, 0.05, label="Force style (IP-Adapter Style)")
-
-            steps  = gr.Slider(10, 60, 30, 1, label="Steps")
-            cfg    = gr.Slider(0.1, 12.0, 6.5, 0.1, label="CFG")
-            width  = gr.Dropdown(choices=[576, 640, 704, 768, 896], value=704, label="Largeur")
-            height = gr.Dropdown(choices=[704, 768, 896, 1024], value=896, label="Hauteur")
-            seed   = gr.Number(value=-1, label="Seed (-1 aléatoire)")
-            btn    = gr.Button("🎨 Générer", variant="primary")
-
-        with gr.Column():
-            out_image = gr.Image(label="Résultat", interactive=False)
-            err_box   = gr.Textbox(label="Erreurs", visible=False)
-            log_box   = gr.Textbox(label="Logs", value="\n".join(load_logs), lines=12)
-
-    def wrap(*args):
-        img, err, logs = generate(*args)
-        return img, gr.update(visible=bool(err), value=err), gr.update(value=logs)
-
-    btn.click(
-        wrap,
-        inputs=[face_image, style_image, prompt, negative,
-                identity_strength, adapter_strength, style_strength,
-                steps, cfg, width, height, seed],
-        outputs=[out_image, err_box, log_box],
-    )
-
-demo.queue(api_open=False)
-if __name__ == "__main__":
-    demo.launch(server_name="0.0.0.0", server_port=7860, ssr_mode=False)
+        load_logs.append("✅ IP-Adapter Style (SDXL) chargé (adapter_na