Increased StyleCLIP video length. Added more descriptions.

README.md

@@ -4,7 +4,7 @@ emoji: 🌖
 colorFrom: blue
 colorTo: gray
 sdk: gradio
-sdk_version: 3.0.1
+sdk_version: 3.0.2
 app_file: app.py
 pinned: false
 license: mit

app.py

@@ -307,14 +307,17 @@ blocks = gr.Blocks()
 with blocks:
     gr.Markdown("<h1><center>StyleGAN-NADA</center></h1>")
     gr.Markdown(
-        "Demo for StyleGAN-NADA: CLIP-Guided Domain Adaptation of Image Generators (SIGGRAPH 2022)."
+        "<h4 style='font-size: 110%;margin-top:.5em'>Inference demo for StyleGAN-NADA: CLIP-Guided Domain Adaptation of Image Generators (SIGGRAPH 2022).</h4>"
     )
     gr.Markdown(
-        "For more information about the paper and code for training your own models (with examples OR text), see below."
+        "<h4 style='font-size: 110%;margin-top:.5em'>Usage</h4><div>Upload an image of your face, pick your desired output styles, and apply StyleGAN-based editing.</div>"
+        "<div>Choose the edit image tab to create static images in all chosen styles. Choose the video tab in order to interpolate between all chosen styles (But take it easy on the servers! We've limited video length, so if you add too many styles, they'll pass in the blink of an eye! 🤗)</div>"
+    )
+    gr.Markdown(
+        "For more information about the paper and code for training your own models (with examples OR text), please visit our <a href='https://stylegan-nada.github.io/' target='_blank'>Project Page</a> or the <a href='https://github.com/rinongal/StyleGAN-nada' target='_blank'>official repository</a>."
     )
 
-
-    gr.Markdown("<h4 style='font-size: 110%;margin-top:.5em'>On biases</h4><div>This model relies on StyleGAN and CLIP, both of which are prone to biases such as poor representation of minorities or reinforcement of societal biases, such as gender norms. </div>")
+    gr.Markdown("<h4 style='font-size: 110%;margin-top:.5em'>A note on social impact</h4><div>This model relies on StyleGAN and CLIP, both of which are prone to biases inherited from their training data and their architecture. These may include (but are not limited to) poor representation of minorities or the perpetution of societal biases, such as gender norms. Moreover, generative models can, and have been used to create deep fake imagery which may assist in the spread of propaganda. However, <a href='https://github.com/NVlabs/stylegan3-detector' target='_blank'>tools are available</a> for identifying StyleGAN generated imagery, and any 'realistic' results produced by this model should be easily identifiable through such tools.</div>")
     
     with gr.Row():
         input_img = gr.inputs.Image(type="filepath", label="Input image")
@@ -330,6 +333,8 @@ with blocks:
                     gr.Markdown("If multiple options are provided, they will be used randomly between images (or sequentially for a video), <u>not</u> together.")
                     gr.Markdown("Please note that some directions may be entangled. For example, hair length adjustments are likely to also modify the perceived gender.")
 
+                    gr.Markdown("For more information about InterFaceGAN, please visit <a href='https://github.com/genforce/interfacegan' target='_blank'>the official repository</a>")
+
                     pose_slider   = gr.Slider(label="Pose", minimum=-1, maximum=1, value=0, step=0.05)
                     smile_slider  = gr.Slider(label="Smile", minimum=-1, maximum=1, value=0, step=0.05)
                     gender_slider = gr.Slider(label="Perceived Gender", minimum=-1, maximum=1, value=0, step=0.05)
@@ -339,8 +344,9 @@ with blocks:
                     ig_edit_choices = [pose_slider, smile_slider, gender_slider, age_slider, hair_slider]
 
                 with gr.TabItem("StyleCLIP Editing Options"):
-                    gr.Markdown("Move the sliders to make the chosen attribute stronger (e.g. the person older) or leave at 0 to disable editing.")
-                    gr.Markdown("If multiple options are provided, they will be used randomly between images (or sequentially for a video), <u>not</u> together")
+                    gr.Markdown("Choose source and target descriptors, such as 'face with hair' to 'face with curly hair'")
+                    gr.Markdown("Editing strength controls the magnitude of change. Disentanglement thresholds limits the number of channels the network can modify, reducing possible leak into other attributes. Setting the threshold too high may lead to no available channels. If you see an error, lower the threshold and try again.")
+                    gr.Markdown("For more information about StyleCLIP, please visit <a href='https://github.com/orpatashnik/StyleCLIP' target='_blank'>the official repository</a>")
 
                     src_text_styleclip = gr.Textbox(label="Source text")
                     tar_text_styleclip = gr.Textbox(label="Target text")
@@ -368,7 +374,7 @@ with blocks:
                     with gr.Row():
                         with gr.Column():
                             gr.Markdown("Warning: Videos generation requires the synthesis of hundreds of frames and is expected to take several minutes.")
-                            gr.Markdown("To reduce queue times, we significantly reduced the number of video frames. Using more than 3 styles will further reduce the frames per style, leading to quicker transitions. For better control, we reccomend cloning the gradio app, adjusting `num_alphas` in `generate_videos`, and running the code locally.")
+                            gr.Markdown("To reduce queue times, we significantly reduced the number of video frames. Using more than 3 styles will further reduce the frames per style, leading to quicker transitions. For better control, we recommend cloning the gradio app, adjusting <b>num_alphas</b> in <b>generate_videos.py</b>, and running the code locally.")
                 with gr.Column():
                     vid_output = gr.outputs.Video(label="Output Video")
 

generate_videos.py

@@ -74,7 +74,7 @@ def generate_frames(source_latent, target_latents, g_ema_list, output_dir):
 
     np_target_latents = [target_latent.cpu().detach().numpy() for target_latent in target_latents]
 
-    num_alphas = min(10, 30 // len(target_latents))
+    num_alphas = 20 if code_is_s else min(10, 30 // len(target_latents))
 
     alphas = np.linspace(0, 1, num=num_alphas)