pjit-21k

pjit-21k

Linum's experimental 2B text to image model, trained for about $21K of compute, that works best with dense 50 to 80 word prompts and cannot write text in images. Suits researchers and curious creators.

Made by Linum-AI on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: pjit-21k, also called P-JiT 21K, is a text to image model from Linum, trained for about $21K of compute. It is a 2B parameter diffusion transformer paired with a frozen Qwen3.5-4B text encoder that downloads separately. Linum calls it a research checkpoint, not a production model, and warns of rough edges. It saw 660M training samples, then a post-training step.

Who it suits: Researchers who study efficient training of image models, and creators who want to try a small experimental model with descriptive prompts.

Industry | Who uses it | How they use it | Example
Research | Image model researchers | Study a low cost training result | Read the code repository and reproduce a sample
Creative work | Illustrators | Make pictures from detailed descriptions | Prompt a children's book scene with a bear and a rabbit
Software | Developers | Test a small text to image model | Run the generate script with a prompt

Skip it if you need words in a picture. It cannot write text, posters, logos or signs, because Linum filtered rendered text out of the training data.

Where to use it:

• A Python environment with the project's code repository, which has a generate script
• OpenRouter, only if you use the optional prompt upsampler

Use case | Who | What they get back
Pictures from dense prompts | Artists | Images made from captions of 50 to 80 words
Pictures from short prompts | Casual users | Images, once the upsampler rewrites the prompt
Training study | Researchers | A small model checkpoint with its code to inspect

How to set it up:

• Python, and the code repository of the same name on GitHub
• Hardware that can run a 2.18B parameter model, 8.7 GB in fp32, plus the separate 4B text encoder
• An OPENROUTER_API_KEY, only if you want the prompt upsampler

1. Get the project's code repository and move into it.
2. Install it in editable mode.
3. Run the generate script with a prompt and a seed.

pip install -e .

pip install -U huggingface_hub
hf download Linum-AI/pjit-21k

How to use it: Write a dense prompt of 50 to 80 words, or let the upsampler expand a short one. Turn the upsampler off with the no upsample flag.

export OPENROUTER_API_KEY=...
python generate.py --prompt "oil painting of weathered fisherman smoking a pipe" --aspect_ratio square --seeds 0

Tool | What it does
generate.py | Makes images from a prompt
--no_upsample | Turns the prompt upsampler off
--aspect_ratio | Picks the frame shape, such as tall or square
--seeds | Sets the random seed
OPENROUTER_API_KEY | The key the upsampler needs

Example: Run the command above with your key set. The upsampler expands the short prompt into a detailed caption, and the model returns a picture of a weathered fisherman smoking a pipe in an oil painting style.

For newcomers and experts: If you are new to this, write one detailed paragraph describing a scene and run it with the no upsample flag, so no outside service is involved.

If you know the field, the base design is P-JiT. Linum trained it for longer, then added a post-training pass using a form of FDFO. Sampling defaults are 50 Euler steps and adaptive projected guidance at scale 15. It was trained at five sizes of about one megapixel.

A dense caption goes in
The frozen Qwen3.5-4B encodes it
The 2B diffusion transformer denoises
An image comes out

Title: Training samples by image size
512px: 610 million
1024px: 50 million

Good to know:

• It struggles with full body shots, many objects at once, several people interacting and world knowledge such as the strings on a guitar.
• The checkpoint file holds the 2.18B parameter transformer without its training only heads.
• License: apache-2.0.
• Cost to run: an OpenRouter key for the optional upsampler, which calls Claude through OpenRouter. The model itself names no cost.
• Data it reaches: with the upsampler, your prompt goes to OpenRouter. The text encoder downloads separately.
• Not for: text, posters, logos or production use, per the card.

On Hugging Face, 13 November 2026

• Task: text to image
• Library: pjit-21k
• Parameters: 2.2B
• Weights: about 8.7 GB in F32; about 1.1 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 623
• Likes: 73
• License: apache-2.0
• Quantized versions: 1
• Fine-tunes: 4
• Tags: pjit-21k, safetensors, text-to-image, diffusion, pixel-space, research

Links

• Model: https://huggingface.co/Linum-AI/pjit-21k
• Files: https://huggingface.co/Linum-AI/pjit-21k/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download Linum-AI/pjit-21k

Then follow the usage in its model card. Nothing in this kit runs by itself.