Image model Jev redraws MaskGIT from first principles by predicting every pixel in parallel

torchcompiled · x · 2026-09-17

Developer anshuc showcased Jev, an image generation model that creates pictures by predicting every pixel in parallel, rather than the usual autoregressive token-by-token approach.

torchcompiled noted this amounts to reinventing MaskGIT from first principles — the masked image generation paradigm Google introduced in 2022, which generates images quickly via parallel prediction of masked tokens. The observation highlights how some newer image models are returning to parallel decoding.

Original post →

More from Multimodal

Multimodal channel →