Update app.py

app.py

@@ -1,9 +1,10 @@
 import sys, os
 sys.path.append("../")
 
-# --- anti-fragmentation VRAM : à définir AVANT toute init CUDA / modèles ---
+# ↓↓↓ ajoute ceci tout de suite après
 os.environ.setdefault("PYTORCH_CUDA_ALLOC_CONF", "expandable_segments:True")
 
+
 import spaces
 import torch
 import random
@@ -36,24 +37,23 @@ ghibli_style_lora_path = hf_hub_download("InstantX/FLUX.1-dev-LoRA-Ghibli", "ghi
 onepiece_style_lora_path = os.path.join(os.path.dirname(__file__), "onepiece_flux_v2.safetensors")
 ONEPIECE_TRIGGER = "onepiece style"
 
+# ---- Universal prompt (homme ou femme)
+UNIVERSAL_PROMPT = (
+    "Upper-body anime portrait of a pirate character inspired by One Piece, confident and charismatic expression, "
+    "original and dynamic pose, expressive eyes with anime-style lighting, slightly windswept hair, preserving the subject’s "
+    "distinctive facial features and hairstyle (and facial hair if present), detailed anime rendering of the face, natural matte skin tone, "
+    "lips matching the skin color (no pink or gloss), wearing stylish pirate clothing appropriate to the subject (open shirt, coat, vest, "
+    "belts, scarves, cape, etc...), with optional pirate accessories (earrings, necklace, bandana or hat) only if they fit the subject’s style, "
+    "well-framed head and shoulders, centered and balanced, cinematic warm lighting, high-quality cel-shaded coloring and clean linework, "
+    "One Piece-style background (ship deck or ocean sky), designed to look cool, original and iconic like a real One Piece portrait character, "
+    "no frame, no text."
+)
+
 # --------------------------------------------
 # Init pipeline
 # --------------------------------------------
 pipe = InstantCharacterFluxPipeline.from_pretrained(base_model, torch_dtype=torch.bfloat16)
 pipe.to(device)
-
-# Offload / slicing / tiling pour réduire les pics VRAM
-try:
-    if hasattr(pipe, "enable_sequential_cpu_offload"):
-        pipe.enable_sequential_cpu_offload()
-    if hasattr(pipe, "vae"):
-        pipe.vae.enable_slicing()
-        pipe.vae.enable_tiling()
-    if hasattr(pipe, "enable_attention_slicing"):
-        pipe.enable_attention_slicing()
-except Exception:
-    pass
-
 pipe.init_adapter(
     image_encoder_path=image_encoder_path,
     image_encoder_2_path=image_encoder_2_path,
@@ -61,10 +61,10 @@ pipe.init_adapter(
 )
 
 # --------------------------------------------
-# Background remover (BiRefNet sur CPU par défaut)
+# Background remover
 # --------------------------------------------
 birefnet = AutoModelForImageSegmentation.from_pretrained(birefnet_path, trust_remote_code=True)
-birefnet.to("cpu")
+birefnet.to(device)
 birefnet.eval()
 birefnet_transform = transforms.Compose([
     transforms.Resize((1024, 1024)),
@@ -74,25 +74,9 @@ birefnet_transform = transforms.Compose([
 
 def remove_bkg(subject_image):
     def infer_matting(img_pil):
-        run_dev = "cuda" if torch.cuda.is_available() else "cpu"
-        try:
-            birefnet.to(run_dev)
-        except Exception:
-            run_dev = "cpu"
-            birefnet.to("cpu")
-
-        inp = birefnet_transform(img_pil).unsqueeze(0).to(run_dev)
+        inp = birefnet_transform(img_pil).unsqueeze(0).to(device)
         with torch.no_grad():
             preds = birefnet(inp)[-1].sigmoid().cpu()
-
-        # libère immédiatement la VRAM
-        try:
-            birefnet.to("cpu")
-        except Exception:
-            pass
-        if torch.cuda.is_available():
-            torch.cuda.empty_cache()
-
         pred = preds[0].squeeze()
         mask = transforms.ToPILImage()(pred).resize(img_pil.size)
         return np.array(mask)[..., None]
@@ -112,11 +96,7 @@ def remove_bkg(subject_image):
     sample_mask = np.stack([mask] * 3, axis=-1)
     obj = sample_mask / 255 * subject_np + (1 - sample_mask / 255) * 255
     cropped = pad_to_square(obj, 255)
-
-    # redimensionne pour se rapprocher du ratio cible (évite étirements)
-    img = Image.fromarray(cropped.astype(np.uint8))
-    img = img.resize((1024, 780), Image.LANCZOS)
-    return img
+    return Image.fromarray(cropped.astype(np.uint8))
 
 # --------------------------------------------
 # Generation logic
@@ -126,10 +106,6 @@ def randomize_seed(seed, randomize):
 
 @spaces.GPU
 def create_image(input_image, prompt, scale, guidance_scale, num_inference_steps, seed, style_mode, negative_prompt=""):
-    # purge VRAM avant
-    if torch.cuda.is_available():
-        torch.cuda.empty_cache()
-
     input_image = remove_bkg(input_image)
 
     if style_mode == "Makoto Shinkai style":
@@ -147,7 +123,7 @@ def create_image(input_image, prompt, scale, guidance_scale, num_inference_steps
         negative_prompt=negative_prompt,
         num_inference_steps=num_inference_steps,
         guidance_scale=guidance_scale,
-        width=1024, height=768,     # <<< sortie fixe 1024 x 768
+        width=1024, height=768,   # <<< sortie fixe 1024 x 768
         subject_image=input_image,
         subject_scale=scale,
         generator=generator,
@@ -157,10 +133,6 @@ def create_image(input_image, prompt, scale, guidance_scale, num_inference_steps
         result = pipe.with_style_lora(lora_file_path=lora_path, trigger=trigger, **common_args)
     else:
         result = pipe(**common_args)
-
-    # purge VRAM après
-    if torch.cuda.is_available():
-        torch.cuda.empty_cache()
     return result.images
 
 # --------------------------------------------
@@ -172,10 +144,11 @@ def generate_fn(image, prompt, scale, style, guidance, steps, seed, randomize, n
 
 title = "🎨 InstantCharacter + One Piece LoRA"
 description = (
-    "Upload your photo, describe your scene, choose **One Piece style**. "
-    "Output is fixed to **1024×768**. API docs enabled (View API) for Make.com."
+    "Upload your photo, use the universal One Piece prompt, choose **One Piece style**. "
+    "Output is fixed to **1024×780**. API is enabled for Make.com."
 )
 
+# (ne PAS mettre api_open ici)
 demo = gr.Interface(
     fn=generate_fn,
     inputs=[
@@ -194,10 +167,10 @@ demo = gr.Interface(
     title=title,
     description=description,
     examples=[
-        ["./assets/girl.jpg", f"a girl playing guitar, {ONEPIECE_TRIGGER}", 0.9, "One Piece style", 3.5, 28, 123, False, ""],
-        ["./assets/boy.jpg",  f"a boy riding a bike, {ONEPIECE_TRIGGER}", 0.9, "One Piece style", 3.5, 28, 123, False, ""]
+        ["./assets/girl.jpg", f"A girl playing guitar, {ONEPIECE_TRIGGER}", 0.9, "One Piece style", 3.5, 28, 123, False, ""],
+        ["./assets/boy.jpg",  f"A boy riding a bike, {ONEPIECE_TRIGGER}", 0.9, "One Piece style", 3.5, 28, 123, False, ""]
     ]
 )
 
-# Affiche la doc API (utile pour trouver l'endpoint /run/predict côté Make)
+# ⇩⇩⇩ utiliser show_api=True ici
 demo.launch(show_api=True)