P-JiT 21K

P-JiT 21K is Linum's (Sahil & Manu Chopra) pixel-space text-to-image research checkpoint: Pyramid-JiT (P-JiT) architecture, ~2B DiT + 4B text encoder, trained on ~660M samples for about $21,038. Weights: https://huggingface.co/Linum-AI/pjit-21k ; code: https://github.com/Linum-AI/pjit-21k ; notes: https://www.linum.ai/field-notes/pjit-21k .

Highlights:

• No VAE — diffusion directly in pixel space with 32×32 patches;
• Explicitly a research artifact, not production; cannot reliably render text/posters;
• Apache-2.0; a stepping stone toward Linum's next video model.

Why it matters

Publishing a reproducible low-cost training ledger with downloadable weights is hard evidence in the T2I cost narrative; pixel-space Pyramid-JiT also differs from mainstream VAE+DiT stacks.

Limits

Struggles with full-body shots, many entities, multi-person interaction, and world knowledge; prompts prefer dense captions (repo ships an upsampler).

1. git clone https://github.com/Linum-AI/pjit-21k && pip install -e .;
2. First run pulls weights and the Qwen3.5-4B text encoder from HF;
3. Needs a CUDA GPU; use --no_upsample to skip prompt expansion;
4. Use under Apache-2.0.