ostris pre-training "Primal VAE": 1-channel latents match modern 8x VAEs via patching
ostrisai · x · 2026-10-26
AI image training tooling author ostris announced he is pre-training a new VAE concept called Primal VAE. It performs only 2x spatial downsampling to a single channel, but patching the latent by 4 yields an effective 8x compression at 16 channels — on par with most modern VAEs — without deep compression. Shared visuals show input vs. reconstruction, the latent visualization, and the difference map, with reconstruction quality looking remarkably close to the original.
Related event: ostris pretrains radical Primal VAE with 1-channel latent(2 posts)→
More from Multimodal
- Dense instance segmentation via Muse Wonder + SAM 3.1 on Meta's Model API — nikhilaravi · 2026-10-26
- LichtFeld Studio v0.5.4.3 ships community-built SteadySplats, POPSpa and Voxelization — janusch_patas · 2026-10-26
- One prompt to a 2-3 min narrated short: hands-on test of Antigravity video generation — AI_Andrew · 2026-10-26
- Creator makes 10,564-frame animated short entirely with Gemini toolchain — AI_Andrew · 2026-10-26
- FLUX.3 Klein step-distilled to 4-8 steps with CFG baked in — Fit-Structure-6489 · 2026-10-26
- Community checkpoint 'Narrative Drift' brings art-directed visuals to Krea 2 Turbo — MoonbearAIArt · 2026-10-26