Janus-Pro: The Open-Source Multimodal Model Challenging DALL-E 3

In January 2025, while global attention was focused on DeepSeek-R1, the DeepSeek-AI team quietly released another revolutionary model: Janus-Pro. Named after the Roman two-faced god Janus, this unified multimodal system can both understand and generate images within a single architecture. It represents a major step forward in open-source multimodal AI.

An Innovative Architecture: Decoupled Visual Encoding

The most distinctive technical feature of Janus-Pro is its decoupled visual encoding. Unlike earlier multimodal models that used the same visual encoder for both understanding and generation, Janus-Pro uses separate pathways for the two tasks.

Specifically, the architecture combines:

• SigLIP-L as the vision encoder for multimodal understanding, supporting 384×384 image inputs;

• A dedicated VQ tokenizer for image generation, with a downsampling rate of 16;

• A single autoregressive transformer that processes both streams, built on DeepSeek-LLM-1.5B-base and 7B-base.

This decoupling reduces the conflict between understanding and generation while increasing the framework’s overall flexibility. The simplicity and effectiveness of the approach make Janus-Pro a strong candidate for the next generation of unified multimodal models.

Record-Breaking Benchmark Performance

Janus-Pro-7B has set new standards in both directions of its capabilities. In multimodal understanding, it achieved a score of 79.2 on MMBench, clearly surpassing previous unified models such as Janus (69.4), TokenFlow (68.9), and MetaMorph (75.2).

In text-to-image generation, Janus-Pro-7B outperformed leading specialized models:

BenchmarkJanus-Pro-7BDALL-E 3Stable Diffusion 3 MediumGenEval0.800.670.74DPG-Bench84.2%––

These results show that Janus-Pro-7B beats DALL-E 3 by 13 percentage points on GenEval and Stable Diffusion 3 Medium by 6 points. Compared with the original Janus, the improvement is even more striking: GenEval rose from 0.61 to 0.80, and DPG-Bench from 79.7% to 84.2%.

Improved Stability and Quality

Janus-Pro also addresses the main limitations of the original Janus. The predecessor suffered from suboptimal performance with short prompts and unstable generation quality. Janus-Pro solves these issues through three dimensions of improvement:

1. Optimized training strategy;

2. Expanded training data, including tens of millions of high-quality synthetic images;

3. Scalability to larger sizes, with 1.5B and 7B parameter versions.

The result is more stable generation for short prompts, improved visual quality, richer details, and the ability to render simple text inside images.

Open Source and Accessibility

Janus-Pro is released under the MIT License for code and the DeepSeek Model License for weights, making it usable even for commercial purposes. The models are available on Hugging Face and GitHub, with community quantizations and GGUF versions for local execution on consumer hardware.

The open-source nature of Janus-Pro represents a direct challenge to OpenAI’s closed-source approach, proving that elegant architecture and efficient training can compete with the brute-force resources of major Western companies.

V67 Janus Pro: The Flux LoRA

The impact of Janus-Pro extends into the creative community as well. The V67 Janus Pro LoRA, available at:

https://civitai.red/models/2611089/v67-janus-pro

is a Flux-based LoRA designed to translate the aesthetics and philosophy of Janus-Pro into a generative visual language. For now, the component is based on Flux, making it compatible with Flux Schnell, Dev, and Pro workflows.

The LoRA is designed to capture the visual identity of the DeepSeek/Janus ecosystem: deep ocean blues, cyan data streams, the iconic whale tail, and a calm, precise aesthetic. It is intended for:

• Conceptual AI art;

• Geopolitical metaphors (USA vs. China AI philosophies);

• Prompt engineering visualization;

• Minimalist, reflective, data-inspired imagery.

The LoRA page provides the full trigger phrase and usage instructions. It represents a bridge between the technical innovation of Janus-Pro and the creativity of the image-generation community.

Conclusion

Janus-Pro is not just another multimodal model: it is a concrete demonstration that algorithmic elegance can prevail over brute computational force. With its decoupled architecture, record-breaking performance, and fully open-source nature, Janus-Pro redefines the standards for unified visual understanding and generation.

Thanks to initiatives such as the V67 Janus Pro LoRA, its influence extends far beyond technical benchmarks and into the territory of visual creativity and artistic expression. The future of multimodal AI speaks Chinese, but above all it speaks the language of accessibility and open innovation.

Angelo Maiota