Jeveration Generates Images in One Pass
- •National University of Singapore researchers introduce Jeveration for image generation and related vision tasks
- •Jeveration generates 512x512 images in 78 ms, up to 122x faster than existing unified models
- •One checkpoint supports text-to-image, image-to-image generation and dense prediction in one forward pass
Researchers Yuhe Liu, Qi Li, Shihua Zhang and Xinchao Wang at the National University of Singapore introduced Jeveration, a model that generates images and performs image-to-image conversion and dense prediction in one forward pass. The paper, published on arXiv and featured as Hugging Face Papers’ #1 Paper of the Day, reports that Jeveration creates a 512x512 image in 78 ms, up to 122x faster than existing unified models, while maintaining competitive image quality and dense prediction performance.
Jeveration repurposes a frozen vision-language model, which processes both images and text, using its visual-semantic knowledge for image generation. The researchers say the challenge is keeping an image coherent when the model predicts all parts in parallel: a text prompt can correspond to many valid images, and the predictions must form a consistent whole. They condition every position on a shared random state so each draw produces a different but coherent image.
Because direct image supervision can lead the model to ignore that random state, the team first learns a distribution over discrete image codes to guide generation in one pass. The same architecture and checkpoint support text-to-image generation, image-to-image generation and dense prediction, a task that assigns a value or label to each image location. The paper says experiments found competitive results for image quality and dense prediction, alongside the reported 78 ms generation time.
The fastest, most reliable AI news, straight to your inbox