Primal VAE released: 80M-param 2x VAE patches to match FLUX 1/2 latents
ostrisai · x · 2026-10-31
Developer ostris open-sourced Primal VAE on Hugging Face in two variants (2x1 and 2x2, compression x channels). It compresses only 2x spatially to 1-2 channels, then DiT-side patch-4 yields latent shapes equivalent to FLUX 1/2 VAEs (16/32 ch @ 8x), beating FLUX 2 VAE on reconstruction. At 80M params, its latent is trained to unit-normal so no scaling/shift factors are needed; images must have sides in multiples of 16. README includes drop-in code for Flux.1/Flux.2 pipelines via diffusers.
Related event: Ostris Open-Sources Primal VAE, Rivals FLUX with Light Compression(2 posts)→
More from Multimodal
- Podcast: solving creator attribution and taste in the age of AI — hsu_byron · 2026-10-31
- fal gives away free Flux 3 Max generations until November 2 — noahsolomon · 2026-10-31
- Genie 4 generates a guy saying "ThursdAI" in real time; Gemini 4 Carbon tops the new index — altryne · 2026-10-31
- Open-source AI image gen stuns with hyper-realistic nine-tailed fox group scene — Some-Dark-5802 · 2026-10-31
- TONGYI SpeechAI: Streaming TTS, not LLMs, is the real bottleneck for voice assistants — rdesh26 · 2026-10-31
- AI-generated 'System Watch' concept with gear-face titanium design fools timeline — Delahuntagram · 2026-10-31