Parallel image generation in one pass: 0.74 Geneval from just 400 GPU-hours on four 96GB GPUs

_akhaliq · x · 2026-11-20

New paper "Jeveration / System One" generates images in a single parallel pass of 128 discrete tokens with bidirectional attention—no iterative denoising or autoregressive decoding—like Janus-Pro but with block decoding instead of token-by-token.

Training cost is remarkably low: starting from Qwen 3.5 4B with LoRA, only four 96GB GPUs and 400 total GPU-hours, reaching 0.74 on Geneval as a unified model. Caveat: standard VLM understanding scores aren't reported, so capability retention there is uncertain.

Related event: Jeveration: Frozen VLM Generates 512px Images in One Forward Pass, 78ms(5 posts)→

Original post →

More from Multimodal

Multimodal channel →