Update app.py

app.py

@@ -166,14 +166,13 @@ def generate(face_image, style_image, prompt, negative_prompt,
         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
+        from PIL import ImageOps
         face = ImageOps.exif_transpose(face_image).convert("RGB")
-        ms = min(face.size)
-        x = (face.width - ms) // 2
-        y = (face.height - ms) // 2
+        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)
+        face_emb, kps_img = extract_face_embed_and_kps(face_sq)  # face_emb: torch.Tensor [1,D] on DEVICE/DTYPE
 
         # IP-Adapter scales
         try:
@@ -189,27 +188,21 @@ def generate(face_image, style_image, prompt, negative_prompt,
         if isinstance(cn, (list, tuple)):
             n_cn = len(cn)
         else:
-            try:
-                n_cn = len(cn)
-            except Exception:
-                n_cn = 1
+            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
-        )
+        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
-        )
+        scale_arg = [scale_val] * n_cn if n_cn > 1 else ([scale_val] if isinstance(cn, (list, tuple)) else scale_val)
 
-        # ✅ Correctif clé : double passage pour compatibilité 0.29 / 0.30
+        # kwargs d’inférence (on met aussi ici pour compat)
         gen_kwargs = dict(
             prompt=(prompt or "").strip(),
             negative_prompt=(negative_prompt or "").strip(),
             image=image_arg,
-            image_embeds=face_emb,
-            added_conditions={"image_embeds": face_emb},     # diffusers ≥ 0.30.0
-            added_cond_kwargs={"image_embeds": face_emb},    # diffusers 0.29.x
+            image_embeds=face_emb,                                # compat pipeline
+            added_conditions={"image_embeds": face_emb},          # diffusers ≥ 0.30.x (si propagé)
+            added_cond_kwargs={"image_embeds": face_emb},         # diffusers 0.29.x (si propagé)
             controlnet_conditioning_scale=scale_arg,
             num_inference_steps=int(steps),
             guidance_scale=float(cfg),
@@ -217,24 +210,46 @@ def generate(face_image, style_image, prompt, negative_prompt,
             height=int(height),
             generator=gen,
         )
-
-        # Si un style est fourni
         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}")
 
-        # Appel pipeline
-        images = pipe(**gen_kwargs).images
+        # 🔧 MONKEY-PATCH: injecter image_embeds au niveau du UNet.forward
+        orig_forward = pipe.unet.forward
+
+        def forward_patch(*args, **kwargs):
+            # on fusionne proprement pour n’écraser rien
+            ac = kwargs.get("added_conditions")
+            if ac is None:
+                ac = {}
+            else:
+                ac = dict(ac)
+            ac["image_embeds"] = face_emb
+            kwargs["added_conditions"] = ac
+            # compat pour 0.29.x
+            kwargs["added_cond_kwargs"] = ac
+            return orig_forward(*args, **kwargs)
+
+        pipe.unet.forward = forward_patch
+
+        try:
+            images = pipe(**gen_kwargs).images
+        finally:
+            # toujours restaurer le forward d'origine
+            pipe.unet.forward = orig_forward
+
         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:
+        import traceback
         return None, "Erreur:\n" + traceback.format_exc(), "\n".join(load_logs)
 
 
+
 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"