tl;dr
DeepSeek, a Chinese AI lab, has released a new family of open-source multimodal AI models called Janus Pro, ranging from 1 billion to 7 billion parameters. The largest version, Janus Pro 7B, reportedly outperforms OpenAI's DALL-E 3 and other leading models on industry benchmarks. The model uses a "n...
DeepSeek AI has released Janus Pro, a family of open-source multimodal AI models that have surpassed industry benchmarks and sparked debate about the future of the AI industry. The largest version, Janus Pro 7B, reportedly outperforms OpenAI's DALL-E 3 and other leading models on industry benchmarks. The model offers a "novel autoregressive framework" and can analyze and generate images at 768x768 resolution. While highly versatile, it may not replace specialized models.
DeepSeek's Janus Pro model uses a "novel autoregressive framework" and is available for immediate download. It ranges from 1 billion to 7 billion parameters, with the largest version, Janus Pro 7B, outperforming OpenAI's DALL-E 3 and other leading models on industry benchmarks GenEval and DPG-Bench. Its release has triggered concerns about its potential to disrupt incumbents and impact major chip manufacturers like Nvidia, which suffered a significant market cap loss following the release. Unlike DeepSeek R1, Janus Pro's full whitepaper is not published, but its technical documentation is available for immediate download for free.
The model's versatility shines in visual understanding, accurately describing elements in a photo and demonstrating good spatial awareness. However, it falls short in tasks requiring reasoning beyond simple descriptions. In image generation, Janus Pro shows robust capabilities but may not match the quality of state-of-the-art models. Although it exceeds in prompt understanding, its execution results in blurry, outdated images in comparison to specialized models. Nevertheless, as an open-source model, Janus's future as a leader among generative AI enthusiasts will depend on updates seeking to improve its capabilities.
Note that there is currently no immediate way to use traditional UIs to run Janus Pro, making it somewhat impractical to run the model locally. Testing spaces have been created by Huggingface users, offering a means to try the model. Users are advised to be cautious of potentially misleading titles in these spaces to ensure accurate and effective testing.