The Janus Pro models have been launched on GitHub and Hugging Face, available in one-billion and seven-billion parameter models. According to DeepSeek, the latter is able to compete with both Stable Diffusion and Dall-E 3 in benchmarking tests, though this hasn’t been thoroughly tested by third pirates right now.

Now DeepSeek has released Janus Pro, an image generating model the company claims that has been engineered for versatility and efficiency. It is and advanced version based on the Janus model, released last year, that performs text-to-image creation that scales up the training data and increases the model size.