Janus: Unified Multi-Modal AI by DeepSeek

The Janus series by DeepSeek offers powerful AI models for unified multimodal understanding and generation. It includes Janus-Pro (advanced reasoning), Janus (decoupled visual encoding), and JanusFlow (harmonized autoregression and rectified flow).

Hey everyone! DeepSeek is ON FIRE! 🔥 They just dropped the Janus series – a new family of AI models focused on unified multimodal understanding and generation. Here's the breakdown: ✨ Janus-Pro: The top-tier model, trained with more data and a larger size for advanced multimodal reasoning and high-quality image generation. 🧩 Janus: Features a decoupled visual encoding architecture, offering flexibility and strong performance in vision-language tasks. ⚡ JanusFlow: Integrates rectified flow with an autoregressive model for enhanced generative capabilities. The Janus series stands out by unifying both understanding and generation across vision and language in a single framework. DeepSeek is also pushing the boundaries with novel architectures, including decoupled visual encoding and the integration of rectified flow. You can download the models now and explore their capabilities!

Looks like @Stable Diffusion has a new competitor... which is free and MIT licensed!

@stable @chrismessina Looks like input is limited to 384 x 384

@stable @chrismessina @rohanrecommends Yes the resolution (both input and output) is the major limitation of these models currently. DeepSeek team is working on that for future versions, as mentioned in their paper.

384x384 input isn't going to do it for me... but let's wait and see.

Love how it gets both aesthetic and functional requirements. Could use better support for CAD file formats though.