demo

demo_tasks/gradio_tasks_relighting.py

@@ -5,7 +5,7 @@ from PIL import Image
 
 
 task_instruction = "Each row shows a process to manipulate the illumination of images and changes the background following the instruction."
-content_instruction = "Beautiful woman, the illumination comes from left side of the image, "
+content_instruction = "In the last row, the illumination comes from left side of the image, with changed background and style as "
 relighting = [
     dict(
         name='sunset over sea', 
@@ -232,7 +232,7 @@ def process_relighting_tasks(x):
             layout_prompt = get_layout_instruction(grid_w, grid_h)
 
             upsampling_noise = 0.6
-            steps = 50
+            steps = 30
             outputs = [mask, grid_h, grid_w, layout_prompt, task_prompt, content_prompt, upsampling_noise, steps] + rets
             break
 

demo_tasks/gradio_tasks_unseen.py

@@ -99,21 +99,25 @@ dense_prediction_data = [
 unseen_tasks = [
     dict(
         name='Frontal Face Reconstruction', 
-        images=[
-            'demo_tasks/examples/face/34e1633a-369f-4324-86c3-3e6418ec00be/face_0.jpg',
-            'demo_tasks/examples/face/34e1633a-369f-4324-86c3-3e6418ec00be/face_2.jpg',
-            'demo_tasks/examples/face/34e1633a-369f-4324-86c3-3e6418ec00be/face_1.jpg',
-            'demo_tasks/examples/face/cb5d403a-f1bb-4392-8302-24846893a797/face_0.jpg',
-            'demo_tasks/examples/face/cb5d403a-f1bb-4392-8302-24846893a797/face_2.jpg',
-            'demo_tasks/examples/face/cb5d403a-f1bb-4392-8302-24846893a797/face_1.jpg',
-            'demo_tasks/examples/face/2ef6aa5a-e751-4bf2-a302-0237ab460627/face_8.jpg',
-            'demo_tasks/examples/face/2ef6aa5a-e751-4bf2-a302-0237ab460627/face_6.jpg',
-            'demo_tasks/examples/face/2ef6aa5a-e751-4bf2-a302-0237ab460627/face_1.jpg',
-        ], 
-        grid_h=3,
-        grid_w=3,
-        task_prompt="Each row presents multi-view of a face, given a frontal face reconstruction task that leverages [IMAGE1] a left side of the face and [IMAGE2] a right side of the face, to generate [IMAGE3] a frontal face that faces the center of the lens.", 
-        content_prompt="The content of the last image in the final row is: the woman's frontal face that faces the center of the lens.", 
+        examples=[
+            dict(
+                images=[
+                    'demo_tasks/examples/face/34e1633a-369f-4324-86c3-3e6418ec00be/face_0.jpg',
+                    'demo_tasks/examples/face/34e1633a-369f-4324-86c3-3e6418ec00be/face_2.jpg',
+                    'demo_tasks/examples/face/34e1633a-369f-4324-86c3-3e6418ec00be/face_1.jpg',
+                    'demo_tasks/examples/face/cb5d403a-f1bb-4392-8302-24846893a797/face_0.jpg',
+                    'demo_tasks/examples/face/cb5d403a-f1bb-4392-8302-24846893a797/face_2.jpg',
+                    'demo_tasks/examples/face/cb5d403a-f1bb-4392-8302-24846893a797/face_1.jpg',
+                    'demo_tasks/examples/face/2ef6aa5a-e751-4bf2-a302-0237ab460627/face_8.jpg',
+                    'demo_tasks/examples/face/2ef6aa5a-e751-4bf2-a302-0237ab460627/face_6.jpg',
+                    'demo_tasks/examples/face/2ef6aa5a-e751-4bf2-a302-0237ab460627/face_1.jpg',
+                ], 
+                grid_h=3,
+                grid_w=3,
+                task_prompt="Each row presents multi-view of a face, given a frontal face reconstruction task that leverages [IMAGE1] a left side of the face and [IMAGE2] a right side of the face, to generate [IMAGE3] a frontal face that faces the center of the lens.", 
+                content_prompt="The content of the last image in the final row is: the woman's frontal face that faces the center of the lens.", 
+            )
+        ],
     ),
     dict(
         name='Image to Depth + Normal + Hed', 

visualcloze.py

@@ -171,9 +171,9 @@ class VisualClozeModel:
             new_w = int(new_h * aspect_ratio)
             target_size = (new_w, new_h)
 
-        if target_size[0] * target_size[1] > 1600 * 1600:
+        if target_size[0] * target_size[1] > 1024 * 1024:
             aspect_ratio = target_size[0] / target_size[1]
-            target_area = 1600 * 1600
+            target_area = 1024 * 1024
             new_h = int((target_area / aspect_ratio) ** 0.5)
             new_w = int(new_h * aspect_ratio)
             target_size = (new_w, new_h)