Upload folder using huggingface_hub

README.md

@@ -1,13 +1,28 @@
 ---
 title: ToDo
+emoji: 🔥
 app_file: app.py
 sdk: gradio
 sdk_version: 4.19.2
 ---
-# ImprovedTokenMerge
+# ToDo: Token Downsampling for Efficient Generation of High-Resolution Images
+---
+
+This is a demo for our recently proposed method, ["ToDo: Token Downsampling for Efficient Generation of High-Resolution Images"](https://arxiv.org/abs/2402.13573), compared against a popular token merging method, ToMe. 
+
+```
+@misc{smith2024todo,
+      title={ToDo: Token Downsampling for Efficient Generation of High-Resolution Images}, 
+      author={Ethan Smith and Nayan Saxena and Aninda Saha},
+      year={2024},
+      eprint={2402.13573},
+      archivePrefix={arXiv}
+}
+```
+
 ![GEuoFn1bMAABQqD](https://github.com/ethansmith2000/ImprovedTokenMerge/assets/98723285/82e03423-81e6-47da-afa4-9c1b2c1c4aeb)
 
-twitter thread explanation: https://twitter.com/Ethan_smith_20/status/1750533558509433137
+blog post: https://sweet-hall-e72.notion.site/ToDo-Token-Downsampling-for-Efficient-Generation-of-High-Resolution-Images-b41be1ac8ddc46be8cd687e67dee2d84?pvs=4
 
 heavily inspired by https://github.com/dbolya/tomesd by @dbolya, a big thanks to the original authors.
 
@@ -15,7 +30,6 @@ This project aims to adress some of the shortcomings of Token Merging for Stable
 I found with the original that you would have to use a high merging ratio to get really any speedups at all, and by then quality was tarnished. Benchmarks here: https://github.com/dbolya/tomesd/issues/19#issuecomment-1507593483
 
 
-
 I propose two changes to the original to solve this.
 1. Merging Method
    - the original calculates a similarity matrix of the input tokens and merges those with highest similarity

app.py

@@ -8,6 +8,15 @@ import math
 import numpy as np
 from PIL import Image
 
+# Globals
+css = """
+h1 {
+  text-align: center;
+  display: block;
+}
+"""
+
+# Pipeline
 pipe = diffusers.StableDiffusionPipeline.from_pretrained("Lykon/DreamShaper").to("cuda", torch.float16)
 pipe.scheduler = diffusers.EulerDiscreteScheduler.from_config(pipe.scheduler.config)
 pipe.safety_checker = None
@@ -70,8 +79,10 @@ def generate(prompt, seed, steps, height_width, negative_prompt, guidance_scale,
 
     return base_img, merged_img, result
 
-with gr.Blocks() as demo:
-    gr.Label("ToDo: Token Downsampling for Efficient Generation of High-Resolution Images")
+
+
+with gr.Blocks(css=css) as demo:
+    gr.Markdown("# ToDo: Token Downsampling for Efficient Generation of High-Resolution Images")
     prompt = gr.Textbox(interactive=True, label="prompt")
     negative_prompt = gr.Textbox(interactive=True, label="negative_prompt")