Nucleus-Image

Nucleus-Image is a text-to-image model built around sparse mixture-of-experts routing. In plain terms, it has a large set of expert blocks, but each generation step only uses the parts it needs instead of running the whole network every time.

That matters if you care about the tradeoff between image quality and inference cost. You still write a normal prompt and get an image back, but the model is designed around using less active compute than a dense model of the same total size.

For example, you can prompt for a product mockup, character portrait, poster-style visual, architectural scene, or wide cinematic frame, then choose one of the supported aspect ratios instead of being locked into a square image.

Start with the Hugging Face Space. Use one prompt where layout matters, such as a product scene with text, objects in specific positions, or a wide environment with foreground and background details. The first thing to check is whether the image follows the spatial request, not only whether it looks polished.

If the browser demo gives you a useful result, move to the model page for local testing through diffusers. The quick start loads NucleusAI/Nucleus-Image, enables the text KV cache, and runs on CUDA with bfloat16. Treat it as a real GPU workflow rather than a lightweight laptop model.

The open inference path is the strongest part of the release. The linked GitHub training recipe appears early, so judge it as a practical text-to-image model first rather than a fully reproducible training release.

• Generate campaign mockups, concept art, product scenes, and editorial visuals from text prompts.
• Test wide, vertical, and standard layouts without changing tools.
• Compare sparse-MoE image generation against dense open image models on the same prompts.
• Prototype prompt sets where object placement, style, and composition matter.
• Use the Space as a fast first pass before deciding whether local GPU testing is worth it.

Try the demo

View model page

Neural Expedition · Useful open-source AI, curated without hype.