cleanup

README.md

@@ -1,21 +1,24 @@
----
-title: Face Editor
-emoji: 🪞
-colorFrom: yellow
-colorTo: indigo
-sdk: gradio
-sdk_version: 3.14.0
-app_file: app.py
-pinned: false
----
-
 # Face Editor
-This face editor uses a CelebA pretrained VQGAN with CLIP to allow prompt-based image manipulation, as well as slider based manipulation using extracted latent vectors. 
 
-I've written a series of Medium articles which provide a detailed and beginner-friendly explanation of how this was built. 
+## Examples
 
-## Features:
-Edit masking using custom backpropagation hook 
 
+## Overview
+This interactive GUI face editor uses a CelebA-pretrained VQGAN-CLIP for prompt-based image manipulation, as well as slider based manipulation using extracted latent vectors. 
+
+I've written a series of Medium articles which provide a detailed and beginner-friendly explanation of how this was built and the intuition behind latent image manipulation. 
+
+[Coming Soon]
+
+## Demo
+Clone the repo and run `app.py` or <a href="https://colab.research.google.com/drive/110uAZIRQjQen0rKqcnX_bqUXIahvRsm9?usp=sharing"> open this colab notebook </a> and run all the cells, then click on the link that appears under the final cell. 
+
+## Features:
+- Positive and negative prompts
+- Multiple prompts
+- Local editing using a gradient masked adversarial loss (implemented as custom pytorch backpropagation hooks). The CLIP loss gradients are masked according to the user's selection, and the LPIPS loss gradients are masked with the inverse of the user mask in order to preserve the initial identity of the face and prevent changes outside of the masked zone that otherwise happen due to latent variable entanglement. 
+- Extracted latent vectors for slider-based editing
+- Rewinding through the history of edits, resuming edits from a previous point in the history
+- Creating GIF animations of the editing process 
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+## Animations

edit.py

@@ -55,8 +55,8 @@ if __name__ == "__main__":
     model.to(device)
     blend_paths(
         model,
-        "./test_data/face.jpeg",
-        "./test_data/face2.jpeg",
+        "./test_pics/face.jpeg",
+        "./test_pics/face2.jpeg",
         quantize=False,
         weight=0.5,
     )

img_processing.py

@@ -9,13 +9,7 @@ import torchvision.transforms.functional as TF
 from PIL import Image, ImageDraw, ImageFont
 
 
-def download_image(url):
-    resp = requests.get(url)
-    resp.raise_for_status()
-    return PIL.Image.open(io.BytesIO(resp.content))
-
-
-def preprocess(img, target_image_size=256, map_dalle=False):
+def preprocess(img, target_image_size=256):
     s = min(img.size)
 
     if s < target_image_size:
@@ -59,18 +53,3 @@ def loop_post_process(x):
     x = get_pil(x.squeeze())
     return x.permute(2, 0, 1).unsqueeze(0)
 
-
-def stack_reconstructions(input, x0, x1, x2, x3, titles=[]):
-    assert input.size == x1.size == x2.size == x3.size
-    w, h = input.size[0], input.size[1]
-    img = Image.new("RGB", (5 * w, h))
-    img.paste(input, (0, 0))
-    img.paste(x0, (1 * w, 0))
-    img.paste(x1, (2 * w, 0))
-    img.paste(x2, (3 * w, 0))
-    img.paste(x3, (4 * w, 0))
-    for i, title in enumerate(titles):
-        ImageDraw.Draw(img).text(
-            (i * w, 0), f"{title}", (255, 255, 255), font=font
-        )  # coordinates, text, color, font
-    return img