The generative AI (GenAI) genie is out of the bottle, and developers of closed models now confront new challenges in monetizing reasoning systems and image generation tools—the latter already widely accessible at low or no cost.
The Chinese startup DeepSeek, once a niche player, has launched Janus-Pro, an open-source multimodal AI that reportedly rivals mainstream image generation apps like OpenAI’s DALL-E 3 and Stable Diffusion in creative tasks such as text-to-image synthesis. Available under an MIT license, Janus-Pro permits free modification and distribution, requiring only attribution and a disclaimer about warranties. By contrast, DALL-E has a closed-source framework, while Stable Diffusion’s licensing terms vary by version.
Janus-Pro follows DeepSeek’s recent release of the open-source R1 reasoning model, which the U.S. AI search firm Perplexity has already integrated into its platform for advanced reasoning tasks. Together, these releases underscore DeepSeek’s aggressive approach to democratize AI tools while undercutting proprietary rivals on cost and transparency.
On Monday, NVIDIA’s stock plummeted 12% as investors grappled with the implications of cheaper, equally potent approaches to genAI model training.
The latest iteration of Janus-Pro builds on the foundational architecture of its predecessor with several upgrades. By decoupling visual encoding pathways—using a SigLIP encoder for understanding tasks (e.g., object recognition) and a VQ tokenizer for generation tasks (e.g., text-to-image synthesis)—the model avoids performance trade-offs that plague unified frameworks. This architecture enables simultaneous high-quality interpretation and generation, a dual capability that sets Janus-Pro apart from DALL-E 3 and Stable Diffusion.
Janus-Pro scores 79.2 on MMBench for multimodal understanding (outperforming MetaMorph and TokenFlow) and 0.80 on GenEval for image generation, surpassing DALL-E 3 (0.67) and Stable Diffusion 3 Medium (0.74).
Nevertheless, Janus-Pro still shares some of its predecessor’s constraints: it caps image outputs at 384×384 pixels, limiting the model’s utility in applications requiring higher resolutions or sharp facial details. Despite these tradeoffs, Janus-Pro’s overall versatility appears to have improved as the paper notes.
