I was also interested in this! 
I tired a few different ways of comparing the diffused images against the original. Comparing the latent embeddings of the SD generated images against the original image looked to work as well or better than the other methods I tried, and doesn’t require another network to embed.
I gave Dreambooth a try, and here are some results using my own face -
Code repo here - Link
For more results, see this notebook - Jupyter Notebook Viewer (nbviewer.org)
Oh wow! Sounds like we more or less took the same direction, and that you did some amazing research here! sorry I missed it (so many amazing works here, it’s hard to keep track of all of them ;)). I will look at your notebook more in depth when I have more time. Thanks!
Guys, I was able to put my DiffEdit version on Gradio spaces, Diffedit - a Hugging Face Space by aayushmnit. Hugging face spaces graciously granted me a free GPU to run my app. Give it a spin
Working through unit 1 of the Diffusion Model class by @johnowhitaker, I created a diffusion model for generating Bored Apes.
Here are a few generations -
I’ve uploaded the dataset and the trained model on HuggingFace Hub.
The notebook for training the model is available on the Github repo.
Maybe late to the party but I was exploring redoing my portfolio website because I will take a sabbatical soon. Tried Github profile README and workflows to pull tweets. I like how it turned out. It’s free, clean, and feature-rich. If anybody wants to build their portfolio website, give Github Profile a try before trying more complex options like Github Pages, Ghost, etc.
I spent that last week diving deeper into this paper ([2211.09794] Null-text Inversion for Editing Real Images using Guided Diffusion Models) that lets you edit images using just text prompts. I tried to explain how it works in my blog post here:
Here are some of the results from playing around with it:
Input Image:
Input Description:
“a chocolate cake”
Edits:
“a red velvet cake”
“a jelly cake”
“a pineapple cake”
This is very impressive, great work!
Thanks tanishq !
Nice! Do you have a tweet I can share?
Thanks Jeremy! Yes:
Spent the weekend implementing the MagicMix: Semantic Mixing with Diffusion Models paper.
The aim of the method is to mix two different concepts in a semantic manner to synthesize a new concept while preserving the spatial layout and geometry.
The method takes an image that provides the layout semantics and a prompt that provides the content semantics for the mixing process.
Here are some examples I reproduced from the paper:
Input Image:
Prompt: “Cake”
Output Image:
Input Image:
Prompt: “Bed”
Output Image:
Input Image:
Prompt: “Family”
Output Image:
Input Image:
Prompt: “Ice-cream”
Output Image:
The method sometimes needs a bit of fiddling around with the parameters to get the best result but overall it was fun implementing the method and reproducing the examples from the paper.
Here is the notebook of the implementation for anyone interested in trying it out.
First preprint! The first in a series of planned papers, explaining visualisations built on the logits from a vision encoder trained using self-supervised learning. The challenge/dataset was this one, and the project repo is here. Here is the paper:
Wow congrats?
I created a callback that calculates the fisher information of the weights during training.
Notebook
Next step is to create a regularization that penalizes for the amount of information stored in the weights.
Really cool to see this coming out of a Zindi challenge! Great work
Last weekend, together with my teammate, we won second place in a $50k AI Hackathon organized by AssemblyAI
Our project was a web app that creates a toy story and its illustration for kids based on the photos of their favorite toy. More about the project here and you can try the app here. The app is not in its best form yet but we are planning to work on it and I would be glad if I get some feedback from you guys.
I am very grateful for the opportuniteis that attending fast ai course is providing me so far. Thank you very much @jeremy!
Thank you! excited to finally start publishing writing from this competition
What a wonderful way to write stories! Very creative work, congratulations
