2-person lab Linum trains 2B text-to-image model from scratch for just $21K, Apache 2.0
ycombinator · x · 2026-11-05
Two-person research lab Linum released a 2B Pyramid-JiT text-to-image model trained from scratch for only $21K, with weights under Apache 2.0 — a milestone toward cheaper video models.
- Pyramid-JiT is a pixel-space architecture that converges 11x faster than its latent-diffusion counterpart (published last month)
- Their thesis: attention is O(n²), so aggressive compression shrinks the context window and slashes training/inference cost — key since today's video models are too slow, expensive, and lack controls for iterative storytelling
- The blog includes 249 sample images across categories, pressure tests, full training settings, and limitations
Goal: creative tools that democratize access to animation.
More from Multimodal
- LFM3-Omni-2B ported to run fully offline on iPhone at ~80 tok/s with custom WebGPU kernels — JosephJacks_ · 2026-11-05
- Dot's creative ChatGPT image-gen workflow draws praise: storybook art beyond one-shot prompts — thederbiedone · 2026-11-05
- HeyGen brings Video 1 to mobile: generate AI avatar videos from a text prompt — HeyGen · 2026-11-05
- U-JiT does pixel-space video generation without a VAE, hitting SOTA on Kinetics-600 — LucaAmb · 2026-11-05
- Rendering a 3D character in Blender, then AI-realistic-ing and animating it for 46 credits — Yobrogamer · 2026-11-05
- Kandinsky open-sources S2V 1.0: speech-driven full-body avatar video in 2B and 21B sizes — Crazy-Repeat-2006 · 2026-11-05