Blog ·

P-JiT 21K online: try the text-to-image model trained for $21K

Linum's P-JiT 21K is a 2B pixel-space image model trained for about $21,000. What makes it interesting, how to write prompts for it, its Apache-2.0 license and how to try it in your browser.

P-JiT 21K is a text-to-image model from Linum with an unusual story: it was trained for about $21,000 of compute. It is a 2-billion-parameter diffusion transformer that works directly on pixels, and its photos and illustrations look far better than its budget suggests. The weights are Apache-2.0; on LiveWeights you can try it in your browser, without a 24 GB GPU.

P-JiT 21KLinum · Apache-2.0A 2B text-to-image model trained for about $21K: photoreal scenes and illustrations from detailed descriptions, in pixel space.Try it online →

What makes P-JiT 21K interesting

  • Pixel space: no VAE. The model predicts pixels directly in 32×32 patches, using the Pyramid-JiT architecture.
  • Cheap to train: 610 million samples at 512 px and 50 million at 1024 px, then a short post-training step, for about $21K.
  • Small: 2.18B parameters plus the Qwen3.5-4B text encoder. A 1-megapixel image takes about 10 seconds on our GPU.

Write long prompts

This is the one rule that matters. P-JiT 21K was trained on dense captions of 50 to 80 words, so short prompts give noticeably worse images. Describe the subject, the setting, the light, the colours and the style in two or three sentences, like the examples on the model page. Also note what the authors say it can't do: it cannot write text (no signs, logos or posters), and it struggles with crowds, full-body shots and interactions between several people.

License: can I use it commercially?

P-JiT 21K is released under Apache-2.0. Linum calls it a research checkpoint, a stepping stone toward their video model, but the license allows commercial use. Your images are yours.

How to try it online

  1. Open the P-JiT 21K page and sign in, then pick a small credit pack (no subscription).
  2. Start from one of the examples (panda, children's book, portrait) and change the details.
  3. Pick a format: square, 4:3, portrait or 16:9, the sizes the model was trained at.
  4. Press Generate. Keep the seed to try small changes to the same picture.

Need text in the image? Use NitroFusion2 instead.

Run it locally

Clone Linum-AI/pjit-21k from GitHub, run pip install -e . and generate.py; the weights (8.7 GB in fp32) and the text encoder (8 GB) download on first use, so plan for a 24 GB GPU. The official script can rewrite short prompts into dense captions using Claude through OpenRouter; with --no_upsample it uses your caption as written.

LiveWeights is not affiliated with the model authors or with the AI Search channel. Facts come from the official pages linked on each model page.