Jeveration Image Model Generates a 512x512 Picture in 78 Milliseconds, Researchers Say
The model handles text-to-image, image-to-image transformation and dense prediction with one set of weights. Its authors claim speeds up to 122 times faster than existing unified models.
Researchers Yuhe Liu, Qi Li, Shihua Zhang and Xinchao Wang have unveiled Jeveration, an AI model that generates a 512x512 image in 78 milliseconds on a single NVIDIA H200 graphics processor.
The authors say the model is up to 122 times faster than existing unified models. The paper was posted Nov. 18, and the speed figures come from the authors' own experiments.
Jeveration uses one set of weights for three kinds of work: generating images from text, transforming existing images and dense prediction. Image-to-image transformation includes colorization and style transfer. Dense prediction covers tasks such as segmentation and depth estimation.
The model is built on a frozen vision-language backbone, meaning the backbone's weights are not retrained. The researchers draw on its visual and semantic knowledge to generate a whole image in a single pass.
To keep results consistent, every position in the image is conditioned on one shared random state. This keeps parallel predictions globally coherent, so each prompt yields a coherent image while outputs can still vary. The authors also introduce a distribution over discrete image codes to guide the single-pass generation.
The researchers say Jeveration keeps competitive image quality and dense-prediction performance while sharply cutting computation time.
The implementation has been released on GitHub with inference code, pretrained weights and a demo. The repository supports Python 3.12 and CUDA-capable GPUs.
The speed numbers are not directly comparable with other recent image-model results. TokenPost's earlier report on FLUX 3 Image Klein speedups used a single NVIDIA B200, 2,048-by-2,048 output and an uncompiled BF16 inference baseline. Jeveration's figures rest on a different GPU, a smaller image size and a different comparison.