Update app.py

app.py

@@ -96,8 +96,8 @@ def generate_caption(image):
         task_prompt = "<MORE_DETAILED_CAPTION>"
         prompt = task_prompt
 
-        inputs = florence_processor(text=prompt, images=image, return_tensors="pt").to(device)
-        
+        inputs = florence_processor(text=prompt, images=image, return_tensors="pt").to(florence_model.device)  # Fixed: Use model's current device instead of static 'device'
+
         generated_ids = florence_model.generate(
             input_ids=inputs["input_ids"],
             pixel_values=inputs["pixel_values"],
@@ -133,8 +133,8 @@ def process_input(input_image, upscale_factor):
         )
         target_input_pixels = MAX_PIXEL_BUDGET / (upscale_factor ** 2)
         scale = (target_input_pixels / (w * h)) ** 0.5
-        new_w = int(w * scale) - int(w * scale) % 8
-        new_h = int(h * scale) - int(h * scale) % 8
+        new_w = int(w * scale) - int(w * scale) % 16  # Fixed: Use % 16 for FLUX alignment (was % 8)
+        new_h = int(h * scale) - int(h * scale) % 16  # Fixed: Use % 16 for FLUX alignment (was % 8)
         input_image = input_image.resize((new_w, new_h), resample=Image.LANCZOS)
         was_resized = True
 
@@ -184,6 +184,9 @@ def tiled_flux_img2img(pipe, prompt, image, strength, steps, guidance, generator
                 generator=generator,
             ).images[0]
 
+            # Fixed: Resize generated tile back to exact tile dimensions if pipeline auto-resized for multiple-of-16 requirement
+            gen_tile = gen_tile.resize((tile_w, tile_h), resample=Image.LANCZOS)
+
             # Paste with blending if overlap
             if overlap > 0:
                 paste_box = (x, y, x + tile_w, y + tile_h)