[Hangzhi]

Adds a DiffusionGemma application at the end of Demos and Applications.

I maintain Diffusion Jev, which serves Google's unchanged DiffusionGemma checkpoint as a Jev-style decision engine with SGLang.

Try the playground: draw a doodle, choose a flower photo, or ask the model to pick an emoji for a message. The app returns a typed answer and candidate scores. Recorded sketch demo.

1. Accuracy close to published Jev results, with lower measured local latency. On the same 231 public JevBench questions, local DiffusionGemma scored 84.8% (196/231), versus 86.6% (200/231) for published Jev 1.13.0 results. Median latency was 183 ms locally versus 665 ms for the published Jev API run. The local test used a warm A100 80GB and loopback requests; the Jev figures include a hosted API network path. These measurements compare deployments with different hardware and network conditions, and do not establish a model-only speed advantage or a full JevBench rank. Report and raw artifacts.
2. An included deployment recipe, with no fine-tuning. The repo provides a pinned SGLang setup, a tested Modal container, and Vercel deployment instructions. The cloud recipe manages dependencies and starts the A100 80GB only when a prediction is requested. It sleeps after idle time; the first prediction may take a few minutes. Deployment guide.
3. Native image input. DiffusionGemma reads the pixels to classify doodles and flowers through the Jev-style decision interface. Jev 1.13 currently accepts text only, according to TypeSafe's model documentation, checked September 22, 2026.

The local API supports choice, noul, and score. This is an independent implementation; candidate scores come from self-conditioned denoising logits and are not calibrated correctness probabilities. Inference code · Typed scoring · Runtime and method.

Tested environment: Linux, A100 80GB, BF16, Python 3.12 and SGLang. The linked demo artifacts contain real inference responses.

Validation: git diff --check; checked the source links and public demo. No duplicate listing was found when submitting.

[google-cla]

Thanks for your pull request! It looks like this may be your first contribution to a Google open source project. Before we can look at your pull request, you'll need to sign a Contributor License Agreement (CLA).

View this failed invocation of the CLA check for more information.

For the most up to date status, view the checks section at the bottom of the pull request.

[Hangzhi]

Close as gemmaDiffusion as Jev is an official support alreay https://x.com/googlegemma/status/2101069861598482817