Buzzing HN · 中文精选· schopra909·· 4 hours agoSelectedAI score68
Linum open-sources 2B text-to-image model Pyramid-JiT, costing roughly $21,000 to train
Show HN: 一个训练成本约为 2 万美元的优质文本转图像模型（Apache 2.0 权重）
AI guide
Linum has open-sourced Pyramid-JiT, a 2B-parameter pixel-space text-to-image model, releasing its weights under the Apache 2.0 license with an end-to-end training cost of just about $21,038. Completely ditching traditional VAEs, the model employs 32×32 token compression to reduce the attention context window by 4x and incorporates multi-resolution intermediate readout heads to accelerate network convergence.
Why it matters
The team validated the feasibility of eliminating traditional VAEs while achieving high-ratio compression directly in pixel space, providing an end-to-end engineering reference for building in-house diffusion models on constrained compute budgets.
Source:Buzzing HN · 中文精选 · linum.ai