Skip to main content
Back to timeline
arXivSource publication:

Jeveration turns a decision model into a one-pass image generator: 78 ms per 512-pixel image, 122 times faster than Janus-Pro-7B

Synopsis

Image generation has long relied on token-by-token autoregression or iterative denoising, requiring many sequential model calls per image. Jeveration treats a whole image as one parallel typed decision: a frozen Qwen3.5-4B vision-language model gains only LoRA adapters and output heads, all code positions share one random state for global consistency, and a learned distribution over discrete image codes guides single-pass generation. On one H200 it produces a 512-pixel image in 78 ms in a single forward pass at 0.74 GenEval, 122 times faster than Janus-Pro-7B, and the same checkpoint also covers image-to-image generation and dense prediction.

Source-provided article image: Jeveration: System One Models Now Generate
arXiv

Interpretation

Image generation gains a route that avoids sequential sampling: the whole image comes back as one set of parallel typed decisions in a single forward pass, conditioned on a text prompt or a source image. Autoregressive models generate token by token and diffusion models denoise iteratively, and both generally need many sequential model evaluations per image, which raises inference latency. On one H200 at batch size 1, Jeveration generates a 512-pixel image end to end in 77.8 ms at 0.74 GenEval, 122 times faster than Janus-Pro-7B in the reported table.

For parallel decisions to form one coherent image, every code position must follow the same source of randomness; otherwise positions can land on incompatible realizations. Letting the model independently pick a codebook entry at each position in parallel cannot capture dependencies between codes, so generated regions may fail to fit together. Without decision training, the first-phase masked model reaches only 0.32 GenEval in one pass; replacing the random state with a fixed one lowers HPSv2.1 from 28.27 to 26.63 and GenEval from 0.73 to 0.70.

One set of shared adapters and task-specific output heads lets a single checkpoint perform text-to-image generation, image-to-image generation and dense prediction, each in one forward pass. Earlier unified multimodal models either generate autoregressively or denoise iteratively, dense prediction usually needs a separate model, and a single image still takes many sequential steps. On NYUv2 segmentation it reaches 67.7 mIoU, on par with dense-only models on the same backbone at 67.8 and 67.3; image-to-image on held-out examples gives 0.880 HED edge F1, 0.802 depth and 0.937 colorization edge F1.

Perspective

The demonstrated scope is text-to-image generation, image-to-image generation and NYUv2 dense prediction at 512-pixel output on one H200 at batch size 1, all from the same Jeveration checkpoint; the material states no next experiment.

Surface-normal accuracy lags dense-only models, which the material attributes to fine local geometry being most sensitive to shared capacity across tasks; image-to-image structure fidelity is measured on 341 held-out examples per task and style PSNR on 223 held-out examples.