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.

DeepSeek says its updated text-to-image generator Janus pro-7B outperforms OpenAI's DALL-E 3 across multiple benchmarks.