Welcome to have a look on our new work:
"Jeveration: System One Models Now Generate"
We started with a fundamental question: Can a model generate an entire image as a single decision?
Jeveration builds on a frozen VLM backbone and chooses all 128 image codes in parallel. Those choices share a random state, so they fit together into one coherent scene. Draw a new state, and the same prompt can produce a different image.
One checkpoint handles all three tasks, each in a single forward pass:
1. Text to image - Generate a tiny dragon asleep in a porcelain teacup.
2. Image to image - Turn an edge map into a realistic scene or a photo into a watercolor.
3. Dense prediction - Segment a scene or estimate its depth from a single photo.
On one H200, Jeveration generates a 512 × 512 image in 78 ms end to end. It scores 0.74 on GenEval and delivers a 122× speedup over Janus-Pro-7B in our evaluation.
Welcome to have a look on our new work: "Jeveration: System One Models Now Generate"
We started with a fundamental question: Can a model generate an entire image as a single decision?
Jeveration builds on a frozen VLM backbone and chooses all 128 image codes in parallel. Those choices share a random state, so they fit together into one coherent scene. Draw a new state, and the same prompt can produce a different image.
One checkpoint handles all three tasks, each in a single forward pass:
1. Text to image - Generate a tiny dragon asleep in a porcelain teacup.
2. Image to image - Turn an edge map into a realistic scene or a photo into a watercolor.
3. Dense prediction - Segment a scene or estimate its depth from a single photo.
On one H200, Jeveration generates a 512 × 512 image in 78 ms end to end. It scores 0.74 on GenEval and delivers a 122× speedup over Janus-Pro-7B in our evaluation.
- Try the demo: https://huggingface.co/spaces/linlinlinlinsdada/Jeveration
- Project page: https://carol-lyh.github.io/Jeveration/
- Paper: https://arxiv.org/abs/2611.24738
- Code and weights: https://github.com/Carol-lyh/Jeveration
Discussions are welcomed, thanks!
We are happy to answer questions and receive feedback!