NUS Releases Jeveration to Generate 512px Images 122x Faster in One Pass
A frozen vision-language model from NUS produces a full 512 pixel image in a single forward pass by treating pixels as one parallel decision.
Takeaways
- Jeveration generates a 512 px image in 78 ms in a single forward pass of a frozen Qwen3.5-4B VLM.
- Reaches 0.74 GenEval, up to 122x faster than Janus-Pro-7B on one H200.
- One checkpoint handles text-to-image, image-to-image, and dense prediction tasks.
- All image codes are predicted in parallel, conditioned on one shared random state for coherence.
- Two-phase training uses a masked model as a reference to guide single-pass decisions.
- Code on GitHub; weights on Hugging Face.
Researchers at the National University of Singapore have released Jeveration, a 5.2-billion-parameter model that predicts the compressed representation of a 512 by 512 image in one transformer pass. The team reports a 0.74 GenEval score and 77.8 ms latency per image on one NVIDIA H200, making the model 122 times faster than Janus-Pro-7B in its comparison. The source code and model weights are publicly available.
One pass replaces the sampling loop
Most image generators rely on sequential computation. Autoregressive models emit tokens one at a time, with each token depending on earlier output. Diffusion models refine noise through repeated denoising steps. Both approaches require several dependent model evaluations before producing an image.
Jeveration represents each image as 128 discrete codes and predicts all of them simultaneously. A separate decoder converts those codes into pixels. The authors describe this output format as a “typed decision,” meaning the model selects a complete output of a requested type, such as an image, segmentation map, or depth map, in parallel.
A frozen base with trainable adapters
The model uses Qwen3.5-4B as its vision-language backbone. Its base weights and vision tower remain frozen while shared rank-256 LoRA adapters and task-specific output heads are trained. About 1.0 billion parameters are trainable, and the complete published model contains 5.2 billion parameters.
A single checkpoint supports text-to-image generation, image-to-image transformation, and dense prediction. Sharing the backbone and adapters reduces the need to maintain separate generative and perception models, although each task still uses an appropriate output head.