GitHub radar
Generate an image with a model trained for just $21,000
P-JiT 21K is a prototype text-to-image model from Linum, a two-brother studio, built to show that a decent model can be trained on a modest budget, not only by big labs. Run it yourself and you get your own image generator, with the full code and weights open for study.
A neural network that turns a text description into a picture. The code and trained model are released openly so anyone can study and reuse them.
When it helps
Worth trying if you want to see how image generation works and run it on your own NVIDIA GPU. For everyday image generation, pick a big commercial generator instead — this one is weaker and can't render text.
Pros
- Code and weights are open — study and reuse them freely
- No paid API needed — the README example works without a key
- Shows a working model can be trained for $21,000, not millions
- Comes with a ready script and example prompt for a first run
Cons
- Can't draw text, signs, logos, or posters
- Needs an NVIDIA GPU with CUDA and plenty of VRAM — no Mac
- Struggles with crowds, full-body shots, and simple logic
How to set it up — step by step
- 1Set up a computer with a CUDA-capable NVIDIA GPU — it won't run on a Mac.
- 2Open your AI agent (Claude Code or Codex) in your project folder.
- 3Send the agent the text from the block below.
- 4Check the resulting image in the outputs folder — ask the agent to show it to you.
- 5If you want to use your own prompts later, get a paid OpenRouter key — not needed for the README example.
Text for your agent
Copy this and send it to your agent — Claude Code, Codex, any of them:
Install P-JiT 21K from https://github.com/Linum-AI/pjit-21k — follow the Getting started steps in the README on a machine with an NVIDIA GPU, generate the portrait from the README's example, and show me the resulting image. Ask me whether you need an OpenRouter key — it's only needed for your own custom prompts, not for the README's example.
You need an NVIDIA GPU with CUDA — it will not run on a Mac or without a GPU. The weights alone are 8.7 GB, plus another 8 GB for the text encoder, so you need a card with real memory headroom.
Open on GitHub▌ More finds