Improve UI

app.py

@@ -26,10 +26,11 @@ _DESCRIPTION = '''
 Given a single-view image and select a target azimuth, ControlNet + SyncDreamer is able to generate the target view
 
 This HF app is modified from [SyncDreamer HF app](https://huggingface.co/spaces/liuyuan-pal/SyncDreamer). The difference is that I added ControlNet on top of SyncDreamer. 
+In addition, the azimuths of both input and output images are both assumed to be 30 degrees.
 '''
 _USER_GUIDE0 = "Step1: Please upload an image in the block above (or choose an example shown in the left)."
-_USER_GUIDE2 = "Step2: Please choose a **Elevation angle** and click **Run Generate**. The **Elevation angle** is the elevation of the input image. This costs about 30s."
-_USER_GUIDE3 = "Generated multiview images are shown below! (You may adjust the **Crop size** and **Elevation angle** to get a better result!)"
+_USER_GUIDE2 = "Step2: Please choose a **Target azimuth** and click **Run Generation**. The **Target azimuth** is the azimuth of the output image relative to the input image in clockwise. This costs about 30s."
+_USER_GUIDE3 = "Generated output image of the target view is shown below! (You may adjust the **Crop size** and **Target azimuth** to get another result!)"
 
 others = '''**Step 1**. Select "Crop size" and click "Crop it". ==> The foreground object is centered and resized. </br>'''
 
@@ -58,7 +59,6 @@ class BackgroundRemoval:
 
     @torch.no_grad()
     def __call__(self, image):
-        # image: [H, W, 3] array in [0, 255].
         image = self.interface([image])[0]
         return image
 
@@ -93,19 +93,16 @@ def sam_predict(predictor, removal, raw_im):
         y_min = int(y_nonzero[0].min())
         x_max = int(x_nonzero[0].max())
         y_max = int(y_nonzero[0].max())
-        # image_nobg.save('./nobg.png')
 
         image_nobg.thumbnail([512, 512], Image.Resampling.LANCZOS)
         image_sam = sam_out_nosave(predictor, image_nobg.convert("RGB"), (x_min, y_min, x_max, y_max))
 
-        # imsave('./mask.png', np.asarray(image_sam)[:,:,3]*255)
         image_sam = np.asarray(image_sam, np.float32) / 255
         out_mask = image_sam[:, :, 3:]
         out_rgb = image_sam[:, :, :3] * out_mask + 1 - out_mask
         out_img = (np.concatenate([out_rgb, out_mask], 2) * 255).astype(np.uint8)
 
         image_sam = Image.fromarray(out_img, mode='RGBA')
-        # image_sam.save('./output.png')
         torch.cuda.empty_cache()
         return image_sam
     else:
@@ -186,16 +183,12 @@ def run_demo():
             with gr.Column(scale=0.8):
                 sam_block = gr.Image(type='pil', image_mode='RGBA', label="SAM output", height=256, interactive=False)
                 crop_size.render()
-                fig1 = gr.Image(value=Image.open('assets/elevation.jpg'), type='pil', image_mode='RGB', height=256, show_label=False, tool=None, interactive=False)
+                fig1 = gr.Image(value=Image.open('assets/azimuth.jpg'), type='pil', image_mode='RGB', height=256, show_label=False, tool=None, interactive=False)
 
             with gr.Column(scale=0.8):
                 input_block = gr.Image(type='pil', image_mode='RGBA', label="Input to SyncDreamer", height=256, interactive=False)
                 azimuth.render()
                 with gr.Accordion('Advanced options', open=False):
-                    cfg_scale = gr.Slider(1.0, 5.0, 2.0, step=0.1, label='Classifier free guidance', interactive=True)
-                    sample_num = gr.Slider(1, 2, 1, step=1, label='Sample num', interactive=False, info='How many instance (16 images per instance)')
-                    sample_steps = gr.Slider(10, 300, 50, step=10, label='Sample steps', interactive=False)
-                    batch_view_num = gr.Slider(1, 16, 16, step=1, label='Batch num', interactive=True)
                     seed = gr.Number(6033, label='Random seed', interactive=True)
                 run_btn = gr.Button('Run generation', variant='primary', interactive=True)
 
@@ -216,8 +209,6 @@ def run_demo():
 
         crop_size.change(fn=resize_inputs, inputs=[sam_block, crop_size], outputs=[input_block], queue=True)\
                  .success(fn=partial(update_guide, _USER_GUIDE2), outputs=[guide_text], queue=False)
-        # crop_btn.click(fn=resize_inputs, inputs=[sam_block, crop_size], outputs=[input_block], queue=False)\
-        #                .success(fn=partial(update_guide, _USER_GUIDE2), outputs=[guide_text], queue=False)
 
         run_btn.click(partial(generate, pipe), inputs=[input_block, azimuth], outputs=[output_block], queue=True)\
                .success(fn=partial(update_guide, _USER_GUIDE3), outputs=[guide_text], queue=False)