Lossless Diffusion : A Friendly Introduction

Lossless Diffusion is a diffusion model that makes files smaller.
It's NOT a data compressor. Rather, an AI model trained to design file formats!
You can try a free encoder here (https://www.fileforma.com/diffusionpng)

Backstory

I took a gap year to start FileForma. We design custom file formats for internet companies.

During COVID, lots of international students were forced to leave the United States and take online classes from their home countries. I was one of them. It was difficult to attend class from my home country, Kenya. The internet was bad! I went down the data compression rabbit-hole and realized two things.

First, data compression is a solved problem. No ✨ magical algorithm ✨ exists to make files infinitely small. Also, you can't compress random data.

Second, understanding your data is the key to smaller file sizes. Data compressors work best when patterns are super evident. That's the reason we trained a diffusion model to help design file formats - AI models are great at finding patterns!

How it was trained

We trained an AI model to design file formats on our behalf. We gave the model a set of .PNG files and asked it to find the best way to pack bytes.

We prompted ChatGPT until we got super interesting C/C++ code. The generated algorithm featured aspects of a diffusion model - mostly the Markob chain - so we called it Lossless Diffusion.

Honestly, ChatGPT appeared to be trained on amazing GitHub repos! We compiled all the amazing patterns we found into a file format here.

The results are super-impressive. My .PNG files are half as small. Interesting enough, pixels don't change. Unlike tools like TinyPNG that change pixel values, our format retains matches exact pixels - before encoding and after decoding. Try it out yourself!

##Feedback
Let me know how it works and reach out in case the website is down. lol
Twitter - https://twitter.com/murage_kibicho
Email - [email protected]