Jeveration Model Generates Images in 78 Milliseconds
Researchers Yuhe Liu, Qi Li, Shihua Zhang and Xinchao Wang introduced a single-pass model that unifies several image tasks with one set of weights.
Updated on Nov. 20, 2026 in Artificial Intelligence

Researchers Yuhe Liu, Qi Li, Shihua Zhang and Xinchao Wang unveiled Jeveration, which generates a 512x512 image in 78 milliseconds on a single NVIDIA H200 graphics processor. The paper was posted November 18, 2026, and the implementation has been released on GitHub.
Why it matters
The model uses one set of weights for text-to-image generation, image-to-image transformation and dense prediction, rather than separate systems for each task. Its speed points to the possibility of faster, lower-cost image tools for developers and users working with visual AI.
The paper reports 78 milliseconds per 512x512 image on a single NVIDIA H200 and a claimed 122-times speedup over existing unified models. Independent verification of these benchmarks and output quality has not been reported.
The details
Jeveration is built on a frozen vision-language backbone whose weights are not retrained, and every position in the image is conditioned on one shared random state. The model draws on the backbone's visual and semantic knowledge to generate a whole image in a single pass, guided by a distribution over discrete image codes. The implementation, including inference code, pretrained weights and a demo, was released on GitHub and supports Python 3.12 and CUDA-capable GPUs.
Timeline
November 18, 2026: Paper posted with the speed figures.
The Tech Race
Jeveration extends the frozen vision-language backbone approach to unify text-to-image generation, image-to-image transformation and dense prediction under one set of weights. The claimed 122-times speedup over existing unified models positions it as a competitor to multi-model image systems in the race for faster generation.
Faster image generation could eventually shorten wait times in creative and design tools that rely on AI. Developers can now access the released code and weights on GitHub to test the model on CUDA-capable GPUs running Python 3.12.
The takeaway
A single model handling several image tasks could simplify how developers build visual AI tools. Readers interested in the approach can review the released implementation on GitHub.
Further reading
Explore more Artificial Intelligence coverage on Wisevoter.
Source note: This article includes information reported by TokenPost.



