Janus

Why we included this project

DeepSeek's Janus series tackles both image understanding and text-to-image generation in one autoregressive model, so you don't have to bolt together separate vision and generation stacks. The trick is splitting visual encoding into separate pathways, letting the same backbone answer questions about an image and also draw new ones from a text prompt without the two tasks fighting over the encoder. For quick experiments, the repo includes ready-to-run inference code for the 1.3B and 7B checkpoints plus a Gradio demo, so you can try captioning, visual question answering, and instruction-following image generation locally before committing to a heavier pipeline. The JanusFlow variant shows how rectified flow can be trained inside a standard LLM framework, which is handy if you want to understand the architecture before adapting it. It's a useful reference point for anyone evaluating open-weight unified multimodal models.