Linum AI releases 2B Pyramid-JiT text-to-image model trained for about $20K with Apache 2.0 weights
- Linum AI published P-JiT 21k, a 2B-parameter Pyramid-JiT text-to-image model, with Apache 2.0 weights on Hugging Face and a reported training cost of about $20K.
- The release page shows a gallery of 249 samples broken down as Stylized 68, Person 57, Product 43, Animal 42, Building 20, and Landscape 19, each clickable for full resolution and its prompt.
- The writeup covers pressure testing the model, the training regime, and a Model Limitations section, with training settings in Appendix A.
- Commenters note the training run used 660M images and prompts with text excluded from the README, and the model still shows common text-to-image failures such as oversmoothing, waxy shine, and mangled hands.
- Discussion points out that the pipeline drops the VAE, which several readers called counterintuitive, and that at INT4 with a swapped text encoder the model could fit in under 2.5GB.
Hacker News 의견들
Wonder how it behaves when quantized. Swap the text encoder for T5-small or CLIP and this could be under 2.5GB at INT4, so we're not that far from running something like this in the browser on WebGPU.
What was the hardest part of this project for you? I'm also curious about efficiency techniques for video gen. What directions look promising right now, and any pointers for a beginner?
For video, a lot of it is efficient attention work, similar to what DeepSeek and Kimi did for LLMs, plus reflow-style tricks to cut down the flow-matching and diffusion iterations. Filtering the CVPR video gen list with an LLM is a decent start.
It looks nice, but it struggles with the same things every t2i model struggles with. Over-smoothing, that waxy shine, mangled limbs.
The big labs are if anything disincentivized to make efficient models, since access to compute is one of their moats. That's why you mostly see efficiency work coming out of Chinese labs. Happy to see a small shop doing this.
That makes no sense. The big labs just don't publish their efficiency research. They can charge the same prices while making models more efficient internally and widen their margins.
What are you using for training data? Which open datasets, or did you pirate it like the big labs?
Even the average finger count looks right, which is rare. Looking at the failures though, I wonder how much cherry picking went into the showcase.
Throwing away the VAE feels counterintuitive and I'm shocked it works that well. Reminds me of that paper where they trained an LLM from scratch that beats LLaMA2-13B for under $0.1M.
Almost every hand is wrong though. Can you tell us more about the 660M training images and prompts, where they came from and what they cover? The README says text was excluded.
The focus on iteration speed instead of just impressive single outputs is refreshing. For animation, revising one part without changing everything else matters a lot. How are you thinking about that control going from images toward video?
If you built an arena where people could drive the cost down to $1K, developers would get really stoked. What's the minimum viable budget for this kind of workload?
This is awesome. Thanks for sharing the weights on HF.