2-person lab trains 2B text-to-image model from scratch for $20K with Pyramid-JiT
generativist · x · 2026-11-05
Two-person lab Linum released Pyramid-JiT, a pixel-space generative architecture they say converges 11x faster than its latent-diffusion counterpart. As a milestone toward cheaper video models, they open-sourced a 2B Pyramid-JiT text-to-image model trained from scratch for only $20K (Apache 2.0 weights on their blog).
Resharing it, angeris commented that if you can train this for $20K, the field hasn't converged on architecture yet — implying today's diffusion-transformer stack may not be the endgame.
More from Multimodal
- 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
- Two photos to a duet video: audio-only reference bypasses Seedance's face filter — Hungry_Incident1146 · 2026-11-05
- Ex-Google founder launches Atmee v2: real-time avatar from 1 photo + 10s audio — alifcoder · 2026-11-05
- Redditor generates detailed fantasy creature locally with Krea2 and MiniMax H3 — -Ghosa- · 2026-11-05
- Topview T1 targets Seedance 2.5-level 1080p at 62% lower cost — alex_verem · 2026-11-05