bug fix

app.py

@@ -27,15 +27,20 @@ vae_model_id = "madebyollin/sdxl-vae-fp16-fix"
 if torch.cuda.is_available():
 
     # load pipe
-    controlnet = ControlNetModel.from_pretrained(controlnet_model_id, variant="fp16", use_safetensors=True, torch_dtype=torch.float16)
-    vae = AutoencoderKL.from_pretrained(vae_model_id, torch_dtype=torch.float16)
+    controlnet = ControlNetModel.from_pretrained(
+        controlnet_model_id, 
+        # variant="fp16",
+        use_safetensors=True,
+        torch_dtype=torch.float32
+    )
+    # vae = AutoencoderKL.from_pretrained(vae_model_id, torch_dtype=torch.float16)
     pipe = StableDiffusionXLControlNetPipeline.from_pretrained(
         base_model_id, 
         controlnet=controlnet, 
-        vae=vae,
-        variant="fp16",
+        # vae=vae,
+        # variant="fp16",
         use_safetensors=True,
-        torch_dtype=torch.float16,
+        torch_dtype=torch.float32,
     )
     pipe.scheduler = EulerDiscreteScheduler.from_config(pipe.scheduler.config)
     pipe.to(device)
@@ -62,7 +67,6 @@ def get_depth_map(image):
     image = feature_extractor(images=image, return_tensors="pt").pixel_values.to("cuda")
     with torch.no_grad(), torch.autocast("cuda"):
         depth_map = depth_estimator(image).predicted_depth
-    print("get_depth_map", original_size)
     depth_map = torch.nn.functional.interpolate(
         depth_map.unsqueeze(1),
         size=original_size,
@@ -93,7 +97,6 @@ def process(image, image_url, prompt, n_prompt, num_steps, guidance_scale, contr
     generator = torch.Generator().manual_seed(seed)
     generated_image = pipe(
         prompt=prompt,
-        image=orginal_image,
         negative_prompt=n_prompt,
         width=size[0],
         height=size[1],
@@ -101,7 +104,7 @@ def process(image, image_url, prompt, n_prompt, num_steps, guidance_scale, contr
         num_inference_steps=num_steps,
         strength=control_strength,
         generator=generator,
-        control_image=depth_image
+        image=depth_image
     ).images[0]
     return [[depth_image, generated_image], "ok"]