ostris Pretrains 'Primal VAE': Aggressive 2x-Downsample, 1-Channel Latent Design
ostrisai · x · 2026-10-26
Researcher ostris is pretraining a new VAE concept, Primal VAE: only 2x spatial downsampling to 1 channel. Patch-4 the latent and you get 8x at 16 channels — same as most modern VAEs — without deep compression.
Current details: pretraining with only LPIPS and MAE for now (critic coming soon); training with a loss to keep the norm in line with noise, plus EQVAE-style latent flips and rotations to make the latent space easier to diffuse.
Related event: ostris pretrains radical Primal VAE with 1-channel latent(2 posts)→
More from Research
- Laying out DeepLIFT as a standalone computation graph: from sequence to attributions directly — jmschreiber91 · 2026-10-26
- Speculative Diffusion Paper Maps the Speed-Quality Pareto Frontier via Two Tradeoffs — UCLA · 2026-10-26
- Git Records Fall Short for Auditing AI's Role in Research Repositories, Paper Finds — Koutian Wu · 2026-10-26
- Fused DeepLIFT explained: why attribution hooks are cheap for Borzoi/AlphaGenome but hurt small models — jmschreiber91 · 2026-10-26
- New paper: AI agents negotiate badly under uncertainty and break deals on purpose — menhguin · 2026-10-26
- Cherimoya v0.3.1 Ships Fused DeepLIFT, ~7x Faster Attribution for Genomics Models — jmschreiber91 · 2026-10-26