P-JiT 21K: A text-to-image research model trained from scratch for 21K USD

Linum's released research checkpoint: P-JiT (Pyramid-JiT), a pixel-space diffusion architecture with a 2B parameter DiT, trained on 6.1 hundred million samples at 512px and 5000 ten thousand samples at 1024px, followed by post-training using an FDFO variant. The authors claim the compute cost was approximately 2.1 ten thousand USD. The text encoder uses a frozen Qwen3.5-4B. Training descriptions are dense captions of 50–80 words, and the code repository includes a script to expand short prompts into this format (requires an OpenRouter API Key, can be disabled). The authors state this is a milestone result for their next-generation video model.

Original title: Linum-AI/pjit-21k

Linum's released research checkpoint: P-JiT (Pyramid-JiT), a pixel-space diffusion architecture with a 2B parameter DiT, trained on 6.1 hundred million samples at 512px and 5000 ten thousand samples at 1024px, followed by post-training using an FDFO variant. The authors claim the compute cost was approximately 2.1 ten thousand USD. The text encoder uses a frozen Qwen3.5-4B. Training descriptions are dense captions of 50–80 words, and the code repository includes a script to expand short prompts into this format (requires an OpenRouter API Key, can be disabled). The authors state this is a milestone result for their next-generation video model.

Why it matters It publicly releases a complete sample of low-cost text-to-image training from scratch, suitable for researchers studying efficient training methods as a reference

Upvotes 66 · Likes 66 · Downloads 382

Reasons

  • Weights, inference code, and a blog post detailing the training approach are all public, Apache-2.0
  • The authors explicitly state the training cost and data scale, facilitating comparisons of training efficiency with other models
  • Prompts and seeds for five samples are provided, allowing for reproduction

Usage

  • After git clone Linum-AI/pjit-21k, run pip install -e . and use generate.py to generate images; defaults are 50 Euler steps and a guidance strength of 15
  • For short prompts, it is recommended to enable prompt expansion, or write detailed descriptions of 50–80 words yourself

Risks

  • The authors explicitly state that the model cannot write text or create posters, as images containing text were intentionally removed during training
  • The model is relatively small, so full-body portraits, multi-person interactions, multiple objects, and common-sense details are prone to errors
  • This is an experimental research checkpoint, not a production model; weights are in fp32, approximately 8.7 GB
  • This entry has not been run; refer to the model card for actual performance

Open

Hugging Face Trending

Explore the latest AI Radar updates