Pillow-rs: Rust port for python pillow (PIL) 12.2.0

[appunni-m]

Hi,

Stackoverflow seems like dead, so I might as well create an issue.
Before this I tried to find any communication detail but couldn't find any

So I have been working on a side project https://github.com/appunni-m/pillow-rs
A Rust port for PIL 1:1 mapping for all API + GPUBatchExecution API for selected operations

I have put in rather decent effort used a lot AI help to get the project to current state. I won't claim it's not SLOP, but it's my slop and I have put atleast 3 months into it before posting here.

I faced majorly 3 issues while building it,

1. The migration and validation itself
2. FreeType
3. Codecs

I built https://github.com/appunni-m/fontdone and https://github.com/appunni-m/image-slash-star to deal with it.

TBH this is a side project, and built on top of PIL, and won't exist without python-pillow. Also I got inspired by Puhu, but puhu was trying to do thread based speedups, which I wanted to avoid and use SIMD based improvements to keep it faster.

It's still on going, agents are going to at improving each operation as we speak. Combined I must have put atleast 100Billion tokens to reach here.

I would like to ask how do I add the project to python-pillow. Wanted to know if there is any scope where I could get help to maintain the project if at all if you think it's worth maintaining.

All inputs are welcome towards the goal. If you need me to prove something ready to do that.

But otherwise you can try it out using

pip install pillow-rs

Also it fully supports WASM as well, so it runs on browser
https://appunni-m.github.io/tiny-image-star/
Created a small demo project to show case that.

Documentation is here : https://appunni-m.github.io/pillow-rs/ if you want to check the benchmarks.

Fontdone is a paraody on fontdue another freetype migration plan, Fontdone is 2x slower than freetype but since it's like 2us per character may not make a whole difference

Image-slash-star is working on getting there on the JPEG encoding benchmarks. Hopefully I can beat JPEG Turbo on it some day which would be a very big achievement according to me.

[appunni-m]

Few things I had added 3 major features which was not there in python-pillow

lazy execution of pipeline - Helps avoid materialization to memory each time instead able to chain SIMD operations, overall helps compiler hinting in LLVM to sequence them. Why I think it could be much faster than the C implementation

explicit GPU batching
Optional multi threading support on operations: pillow-rs[parallel] -> feature flagged in rust and extra package installation in python

https://appunni-m.github.io/pillow-rs/compatibility/#pillow-rs-additions-beyond-pillow

Let me know if you find this interesting and have questions on it

And I have used mostly GPT 5.6 and 6 Luna Max on it, very less Sol and some deepseek models initially. No claude models were used.