Jeveration generates 512px images in 78ms in one forward pass, 122x faster than Janus-Pro
_akhaliq · x · 2026-11-19
New paper Jeveration shows "System One" models can generate too: a frozen vision-language model (Jev) repurposes its visual-semantic knowledge for holistic image generation in a single forward pass.
- The key challenge is global consistency: one prompt admits many valid images, so all positions are conditioned on a shared random state, with a learned distribution over discrete image codes guiding single-pass generation.
- One checkpoint unifies text-to-image, image-to-image (structure control, style transfer) and dense prediction (segmentation/depth/normals), each in one forward pass.
- Speed: 512×512 in 78ms end-to-end on one H200 — 122x faster than Janus-Pro-7B (9.51s) and well ahead of FLUX.1-schnell (0.35s) and SD3-medium (2.09s), with competitive quality.
Related event: Jeveration: Frozen VLM Generates 512px Images in One Forward Pass, 78ms(5 posts)→
More from Multimodal
- Building a Claude-controlled video editor over MCP: lessons from ~200 tools, approvals and context bloat — MrxAnomynous · 2026-11-20
- OpenMontage: a self-hosted pipeline for consistent AI short dramas — Direct_Grab7063 · 2026-11-20
- CapCut launches ChatGPT video plugin: generate and edit timeline-based videos in chat — alifcoder · 2026-11-20
- I gave Claude a video editor — it made a 21-shot short I can fix shot-by-shot — MrxAnomynous · 2026-11-20
- Demi Guo previews real-time AI video model for directing films like a set — ying11231 · 2026-11-20
- Full automated GTA 6 content pipeline: Apify + ElevenLabs + Claude + Codex — eptwts · 2026-11-20