News · 2026-11-19
Jeveration paper tests parallel image-code generation
Researchers at the National University of Singapore have presented Jeveration, a model that selects image codes in parallel in one vision-language model pass before a separate decoder renders the result. Their preprint reports fast image generation against selected baselines, with tradeoffs in quality and limited independent validation.
Key facts
- The preprint was submitted November 18, 2026.
- The model selects 128 codes from an 8,192-entry codebook for a 512-pixel image.
- The authors report 77.8 milliseconds on an H200 for their end-to-end batch-one setup.
- Three released tensor files total about 6.5 GB, with separate dependencies; minimum inference VRAM is unstated.
An image generator can build a result in stages. Jeveration starts from a frozen Qwen3.5-4B vision-language model with adapters and task-specific heads. It chooses a compact set of image codes in parallel, and a decoder turns them into pixels. Imagine a designer selecting a set of numbered pattern pieces at once, then a separate workshop assembling the picture. The one-pass claim describes the selection step, not the entire rendering pipeline.
Parallel decisions can clash: neighboring choices may be individually plausible but inconsistent together. The authors condition all positions on a shared random state and report internal ablations supporting that coordination. The full paper describes a training procedure that first learns masked code distributions and then trains the one-pass model to match them. The phrase “System One” is the authors' framing for fast typed decisions, informed by TypeSafe's terminology, not an established proof of a new general model category.
For a closer comparison, the authors report 0.74 on GenEval and 77.8 milliseconds for Jeveration, versus 0.63 and 94.1 milliseconds for Show-o Turbo-512 at the same listed resolution. Jeveration scores lower on DPG-Bench. Those are paper-reported results, not independent reruns. The more dramatic 122.2-fold latency ratio compares Jeveration with Janus-Pro-7B at a different resolution; Janus has higher scores on two reported quality measures. Calling it 122 times faster at equal quality would go beyond the table.
The project page and inference repository describe generation, image transformations and dense prediction tasks. The model card and weights list three tensor files totaling about 6.5 GB, plus separately downloaded backbone and other components. The authors timed inference on an H200, a GPU with 141 GB of memory; that is the test configuration, not a minimum requirement. The reviewed repository does not include the full training code or packaged datasets.
The paper's title calls the method “System One Models Now Generate.” Its contribution is more precise: a shared-state parallel code selection mechanism plus a decoder and task heads. Hugging Face's paper page shows early interest, while the reviewed sources do not establish an independent reproduction. The next meaningful comparison would hold resolution, quality, hardware and timing method as constant as possible.
The authors say the same checkpoint also supports segmentation, depth and surface-normal prediction, plus several image-to-image transformations. That is a wider task package than text-to-image alone, although the model card limits the dense-head training to indoor NYUv2 imagery and describes English prompts. Performance on those tasks should not be assumed to transfer to arbitrary scenes or languages.
The evaluation's timing protocol is also specific: the paper reports the median of five calls after two warm-up calls, batch size one, on the same H200 for the listed timing comparisons. Some quality figures in the table come from other authors or papers rather than a single uniform rerun. This makes the matched-resolution comparison informative but still conditional on the disclosed settings.
Key questions
Does Jeveration generate pixels in one pass?
Is it 122 times faster at the same quality?
Can people reproduce the training?
Cite this
APA
Ground Truth. (2026, November 19). Jeveration paper tests parallel image-code generation. Ground Truth. https://groundtruth.day/news/jeveration-parallel-image-codes-research.html
BibTeX
@misc{groundtruth:jeveration-parallel-image-codes-research,
title = {Jeveration paper tests parallel image-code generation},
author = {{Ground Truth}},
year = {2026},
month = {nov},
url = {https://groundtruth.day/news/jeveration-parallel-image-codes-research.html}
}
Comments are replies to this story on Bluesky — reply with any Bluesky account to join in.