22

app.py

@@ -113,7 +113,7 @@ def inference(
     num_inference_steps=8,
     progress=gr.Progress(track_tqdm=True),
 ):
-    # guidance_scale=2
+    # guidance_scale=1
     timestamp = time.time()
     print(f"timestamp: {timestamp}")
 
@@ -125,9 +125,10 @@ def inference(
     
     upscale_target = 2
     upscale_nearest = 16
+    upscale_max_size = 1440
     # rescale to prevent OOM
     input_image = edit_dict['background']
-    input_image, width, height = image_utils.rescale_image(input_image, upscale_target, upscale_nearest, max_size=1280)
+    input_image, width, height = image_utils.rescale_image(input_image, upscale_target, upscale_nearest, max_size=upscale_max_size)
     sample_size = [height, width]
     
     print("DEBUG: inpaint_image")
@@ -138,18 +139,18 @@ def inference(
 
     print("DEBUG: mask_image")
     if mask_image is not None:
-        mask_image, w, h = image_utils.rescale_image(mask_image, upscale_target, upscale_nearest, max_size=1280)
+        mask_image, w, h = image_utils.rescale_image(mask_image, upscale_target, upscale_nearest, max_size=upscale_max_size)
         mask_image = get_image_latent(mask_image, sample_size=sample_size)[:, :1, 0]
     else:
         mask_image = torch.ones([1, 1, sample_size[0], sample_size[1]]) * 255
 
-    # print("DEBUG: control_image_torch")
-    # processor = Processor('canny')
-    # control_image, w, h = image_utils.rescale_image(input_image, scale_target, upscale_nearest, max_size=1280)
-    # control_image = control_image.resize((1024, 1024))
-    # control_image = processor(control_image, to_pil=True)
-    # control_image = control_image.resize((width, height))
-    # control_image_torch = get_image_latent(control_image, sample_size=sample_size)[:, :, 0]
+    print("DEBUG: control_image_torch")
+    processor = Processor('openpose_full')
+    control_image, w, h = image_utils.rescale_image(input_image, scale_target, upscale_nearest, max_size=1280)
+    control_image = control_image.resize((1024, 1024))
+    control_image = processor(control_image, to_pil=True)
+    control_image = control_image.resize((width, height))
+    control_image_torch = get_image_latent(control_image, sample_size=sample_size)[:, :, 0]
 
     # generation
     if randomize_seed: seed = random.randint(0, MAX_SEED)
@@ -164,7 +165,7 @@ def inference(
         guidance_scale=guidance_scale,
         image = inpaint_image,
         mask_image = mask_image,
-        # control_image=control_image_torch,
+        control_image=control_image_torch,
         num_inference_steps=num_inference_steps,
         control_context_scale=control_context_scale,
     ).images[0]
@@ -240,7 +241,7 @@ with gr.Blocks() as demo:
                     guidance_scale = gr.Slider(
                         label="Guidance scale",
                         minimum=0.0,
-                        maximum=3.0,
+                        maximum=1.0,
                         step=0.1,
                         value=1.0,
                     )