Update app.py

app.py

@@ -22,8 +22,8 @@ from pipeline import InstantCharacterFluxPipeline
 MAX_SEED = np.iinfo(np.int32).max
 device = "cuda" if torch.cuda.is_available() else "cpu"
 dtype = torch.float16  # L4: FP16 OK
-HF_TOKEN = os.environ.get("HF_TOKEN") or os.environ.get("HUGGINGFACEHUB_API_TOKEN")
 
+HF_TOKEN = os.environ.get("HF_TOKEN") or os.environ.get("HUGGINGFACEHUB_API_TOKEN")
 def need_token_guard():
     if HF_TOKEN is None:
         raise gr.Error(
@@ -36,32 +36,32 @@ def need_token_guard():
 # =========================
 base_model = "black-forest-labs/FLUX.1-dev"
 image_encoder_path = "google/siglip-so400m-patch14-384"
-image_encoder_2_path = "facebook/dinov2-giant"
+# 🔻 On supprime l'encodeur 2 (DINOv2-giant) pour sauver ~10 Go RAM
+image_encoder_2_path = None
 birefnet_path = "ZhengPeng7/BiRefNet"
 
 # Ton LoRA One Piece (local)
-onepiece_flux_lora_path = "./onepiece_flux_v2.safetensors"  # assure-toi qu'il est bien à la racine
+onepiece_flux_lora_path = "./onepiece_flux_v2.safetensors"
 onepiece_flux_trigger = "onepiece style"
 
 def _dl(repo_id, filename, token=None):
     return hf_hub_download(repo_id=repo_id, filename=filename, token=token)
 
 need_token_guard()
-# Uniquement l'IP-Adapter (nécessaire à l'identité)
+# Uniquement l'IP-Adapter (nécessaire à l'identité) — 5.6 Go
 ip_adapter_path = _dl("tencent/InstantCharacter", "instantcharacter_ip-adapter.bin", HF_TOKEN)
 
 # =========================
-# PIPELINE (GPU only, low RAM)
+# PIPELINE (GPU only, low RAM peak)
 # =========================
 pipe = InstantCharacterFluxPipeline.from_pretrained(
     base_model,
     torch_dtype=dtype,
     token=HF_TOKEN,
-    low_cpu_mem_usage=True,   # réduit le pic RAM au chargement
+    low_cpu_mem_usage=True,   # ↓ pic RAM à l'init
 )
 pipe.to(device)
 
-# xFormers si dispo (baisse l'empreinte GPU)
 try:
     pipe.enable_xformers_memory_efficient_attention()
 except Exception:
@@ -72,11 +72,16 @@ if hasattr(pipe, "vae"):
     if hasattr(pipe.vae, "enable_slicing"): pipe.vae.enable_slicing()
     if hasattr(pipe.vae, "enable_tiling"):  pipe.vae.enable_tiling()
 
-pipe.init_adapter(
+# 🔻 Init de l'adapter: 1 seul image encoder (SigLIP) + moins de tokens
+adapter_kwargs = dict(
     image_encoder_path=image_encoder_path,
-    image_encoder_2_path=image_encoder_2_path,
-    subject_ipadapter_cfg=dict(subject_ip_adapter_path=ip_adapter_path, nb_token=1024),
+    subject_ipadapter_cfg=dict(subject_ip_adapter_path=ip_adapter_path, nb_token=512),  # 1024 -> 512 pour baisser mémoire
 )
+# N'ajoute image_encoder_2_path que s'il existe
+if image_encoder_2_path:
+    adapter_kwargs["image_encoder_2_path"] = image_encoder_2_path
+
+pipe.init_adapter(**adapter_kwargs)
 
 # =========================
 # BiRefNet : lazy-load sur CPU
@@ -156,12 +161,10 @@ def create_image(
 ):
     if input_image is None:
         raise gr.Error("Merci d'uploader une image de visage.")
-
     if use_onepiece_lora and not os.path.exists(onepiece_flux_lora_path):
         raise gr.Error(f"Fichier LoRA manquant : {onepiece_flux_lora_path}")
 
     input_image = remove_bkg(input_image)
-
     generator = torch.Generator(device=device).manual_seed(int(seed))
 
     if use_onepiece_lora:
@@ -195,9 +198,9 @@ def create_image(
 # =========================
 # UI
 # =========================
-title = "<h1 align='center'>InstantCharacter (FLUX.1-dev) + One Piece (FLUX LoRA) — minimal</h1>"
+title = "<h1 align='center'>InstantCharacter (FLUX.1-dev) + One Piece (FLUX LoRA) — single encoder</h1>"
 description = (
-    "GPU-only (FP16) pour éviter la limite RAM 30 Go. "
+    "GPU-only (FP16), low_cpu_mem_usage=True, **sans DINOv2-giant** pour éviter la limite RAM 30 Go. "
     "Départ en 768×768, tu peux monter à 896→1024 si stable."
 )
 
@@ -208,12 +211,10 @@ with block:
     with gr.Row():
         with gr.Column():
             image_pil = gr.Image(label="Source Image", type="pil")
-
             prompt = gr.Textbox(
                 label="Prompt",
                 value="onepiece style, a pirate character standing on a ship deck, shonen manga, strong black line art, cel shading, expressive eyes, dynamic pose, clean linework"
             )
-
             scale = gr.Slider(0.0, 1.5, 1.0, 0.01, label="Scale (face strength)")
             use_onepiece_lora = gr.Checkbox(value=True, label="Use One Piece (FLUX LoRA)")
             lora_strength = gr.Slider(0.0, 1.5, 0.85, 0.05, label="LoRA strength")