Update app.py

app.py

@@ -9,7 +9,7 @@ from PIL import Image
 import gradio as gr
 
 from huggingface_hub import hf_hub_download
-from transformers import AutoModelForImageSegmentation
+from transformers import AutoModelForImageSegmentation, CLIPProcessor, CLIPModel
 from torchvision import transforms
 from pipeline import InstantCharacterFluxPipeline
 
@@ -33,6 +33,31 @@ 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"
 
+# ---------- Gender-aware prompt templates ----------
+MALE_PROMPT = (
+    "Upper-body anime portrait of a female pirate character inspired by One Piece, "
+    "confident and charismatic expression, elegant yet powerful pose, playful or determined smirk, "
+    "expressive eyes with anime-style lighting, slightly windswept hair, keeping the distinctive facial features and hairstyle of the person, "
+    "detailed anime rendering of their face, natural matte skin tone, lips matching the skin color (no pink or gloss), "
+    "wearing optional stylish pirate clothing inspired by One Piece (open shirt, corset, belts, jackets, jewelry, or scarves), "
+    "with optional pirate accessories such as earrings, necklace, or hat, only if it fits the character’s style, "
+    "well-framed composition showing the full head and shoulders clearly, centered and balanced, cinematic warm lighting, "
+    "high-quality cel-shaded coloring and detailed linework, One Piece-style background (ship deck or ocean sky), "
+    "designed to look beautiful, elegant and iconic like a real One Piece heroine character, no frame, no text."
+)
+
+FEMALE_PROMPT = (
+    "Upper-body anime portrait of a female pirate character inspired by One Piece, "
+    "confident and charismatic expression, elegant yet powerful pose, playful or determined smirk, "
+    "expressive eyes with anime-style lighting, slightly windswept hair, keeping the distinctive facial features and hairstyle of the person, "
+    "detailed anime rendering of their face, natural matte skin tone, lips matching the skin color (no pink or gloss), "
+    "wearing optional stylish pirate clothing inspired by One Piece (open shirt, corset, belts, jackets, jewelry, or scarves), "
+    "with optional pirate accessories such as earrings, necklace, or hat, only if it fits the character’s style, "
+    "well-framed composition showing the full head and shoulders clearly, centered and balanced, cinematic warm lighting, "
+    "high-quality cel-shaded coloring and detailed linework, One Piece-style background (ship deck or ocean sky), "
+    "designed to look beautiful, elegant and iconic like a real One Piece heroine character, no frame, no text."
+)
+
 # --------------------------------------------
 # Init pipeline
 # --------------------------------------------
@@ -82,6 +107,22 @@ def remove_bkg(subject_image):
     cropped = pad_to_square(obj, 255)
     return Image.fromarray(cropped.astype(np.uint8))
 
+# --------------------------------------------
+# Simple gender detector (CLIP zero-shot)
+# --------------------------------------------
+clip_model = CLIPModel.from_pretrained("openai/clip-vit-base-patch32").to(device)
+clip_processor = CLIPProcessor.from_pretrained("openai/clip-vit-base-patch32")
+clip_model.eval()
+
+@torch.no_grad()
+def detect_gender(img_pil: Image.Image) -> str:
+    texts = ["a portrait photo of a man", "a portrait photo of a woman"]
+    inputs = clip_processor(text=texts, images=img_pil.convert("RGB"), return_tensors="pt", padding=True).to(device)
+    outputs = clip_model(**inputs)
+    logits_per_image = outputs.logits_per_image.squeeze(0)  # (2,)
+    idx = int(torch.argmax(logits_per_image).item())
+    return "male" if idx == 0 else "female"
+
 # --------------------------------------------
 # Generation logic
 # --------------------------------------------
@@ -104,7 +145,7 @@ def create_image(input_image, prompt, scale, guidance_scale, num_inference_steps
     generator = torch.manual_seed(seed)
     common_args = dict(
         prompt=prompt,
-        negative_prompt=negative_prompt,   # <-- ajouté
+        negative_prompt=negative_prompt,
         num_inference_steps=num_inference_steps,
         guidance_scale=guidance_scale,
         width=1024, height=1024,
@@ -122,14 +163,20 @@ def create_image(input_image, prompt, scale, guidance_scale, num_inference_steps
 # --------------------------------------------
 # UI definition (Gradio 5)
 # --------------------------------------------
-def generate_fn(image, prompt, scale, style, guidance, steps, seed, randomize, negative_prompt):
+def generate_fn(image, prompt, scale, style, guidance, steps, seed, randomize, negative_prompt, auto_prompt):
+    # si auto_prompt est activé, on remplace le prompt par le template choisi via CLIP
+    if auto_prompt and image is not None:
+        gender = detect_gender(image)
+        chosen = MALE_PROMPT if gender == "male" else FEMALE_PROMPT
+        # on ajoute le trigger One Piece pour renforcer le style
+        prompt = f"{chosen}, {ONEPIECE_TRIGGER}"
     seed = randomize_seed(seed, randomize)
     return create_image(image, prompt, scale, guidance, steps, seed, style, negative_prompt)
 
 title = "🎨 InstantCharacter + One Piece LoRA"
 description = (
-    "Upload your photo, describe your scene, choose **One Piece style** "
-    "and generate yourself as an anime character!"
+    "Upload your photo, describe your scene, or tick **Auto One Piece Prompt** to auto-pick a gender-aware template. "
+    "Choose **One Piece style** to apply the LoRA."
 )
 
 demo = gr.Interface(
@@ -144,15 +191,16 @@ demo = gr.Interface(
         gr.Slider(5, 50, value=28, step=1, label="Inference Steps"),
         gr.Slider(-1000000, 1000000, value=123456, step=1, label="Seed"),
         gr.Checkbox(value=True, label="Randomize Seed"),
-        gr.Textbox(label="Negative Prompt", placeholder="e.g. photorealistic, realistic skin, pores, hdr")  # <-- ajouté
+        gr.Textbox(label="Negative Prompt", placeholder="e.g. photorealistic, realistic skin, pores, hdr"),
+        gr.Checkbox(value=True, label="Auto One Piece Prompt (gender-aware)"),   # <--- nouvel interrupteur
     ],
     outputs=gr.Gallery(label="Generated Image"),
     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, "", True],
+        ["./assets/boy.jpg",  f"A boy riding a bike, {ONEPIECE_TRIGGER}", 0.9, "One Piece style", 3.5, 28, 123, False, "", True],
+    ],
 )
 
 demo.launch()