So how did I get here? :D
The last two weeks i've been messing around with pdfs. I made slop pdf, a pdf that generates random ai slop stories when you open it, and I was showing that off to one of my friends who suggested I get the LLM to actually work inside the pdf, without an API call to a relay.
I laughed at the idea, but then we did some research together, and we managed to find EvanZhouDev's llm.pdf. Because it existed, I didn't want to waste my time or my tokens building something like it. So I pivoted, eventually deciding to make a diffusion image model work fully inside a pdf.
I forked EvanZhouDev's llm.pdf repo, and eventually managed to get it working (with the help of claude). Here's the finished version.
proof of conceptness disclaimer/warning :p
This model only has 131 Known Objects. If you don't use one of these objects, it will try to map your input to a known object, terribly.
Requirements
- Chrome or any other chromium browser (minus safari) -- plz note no safari/firefox support
- at least 75MB of storage (sorry its such a big file
)
So is this AI?
Ehh.. it is technically a denoising diffusion model similar to Stable Diffusion, Midjourney or Dall-e but like way worse
I trained this model on 130ish categories using Google's Quick Draw dataset of doodle images. It took me so so long (because I did it on my Macbook Air) but eventually I trained it on enough data, for it to produce recognisable stuff.
This training took alot, and ended up inflating the PDF's size to 78mb :cry:.
Now you may be wondering, how'd i get the noise?.
I get the noise by using the current time, legit the only way to get something noisy from a pdf. ;P
Why is it ASCII art?
It was easier to just render the drawing as 28 rows of text, due to the nature of PDF form fields.
You may notice that the actual generation is inverted, this is so that it shows up well against a white background.
So how fast is it?
- It takes about 1.8 seconds per step when I have the PDF generating in Chrome on my M5 Macbook Air
- The PDF has a compute of 0.116 GFLOPS on my my laptop when opened in Chrome (58 Million MAC/s)
les specs
| # of Parameters | 55,403,136 int8 |
|---|---|
| PDF Size | 75 MB |
| Vocabulary Size | 131 doodle categories |
| Compute | 886M MAC per image |
| Speed | ~15.3s per image (at 58M MAC/s) |
| Quality | 0.874 class accuracy |
| Training | 45 epochs, 327,500 drawings, final loss 0.2476 |
roll the credits
Built by me, Alex Van Doren. I'm one of the developers behind Hack Club's Stardance Challenge, I'm a super cool sailor, and I'm also an incoming freshman at Brown University.
This is a fork of the really cool llm.pdf by Evan Zhou, that runs an LLM inside a PDF, the thing I was originally intending to build. Make sure to check out his Youtube Video about it, which honestly explained alot.
I also wanted to credit linuxpdf and DoomPDF by ading2210. They inspired me to experiment within the PDF spec, and the coolest thing about them is that they were actually submitted to a Hack Club program I worked on, The Summer of Making.Also ty Google, for making the Quick, Draw! dataset.
Copyright © 2026 vandorena. Source on GitHub. | Check out my personal site alexvd.dev