DeepSeek surprised the market when its assistant based on the DeepSeek-V3 model reached the top spot in Apple’s App Store in the United States, overshadowing long-standing applications. Now, all eyes are on Janus Pro 7B, the latest iteration of its AI Model specialized in image generation.
In a recent technical report, the company claims that Janus Pro 7B —the most robust version of the family— has outperformed DALL-E 3 from OpenAI and Stable Diffusion from Stability AI in significant benchmark tests.
This news has not gone unnoticed, especially at a time when more industries (marketing, education, entertainment) are seeking to harness AI’s potential for creating high-quality visual content.
According to DeepSeek, Janus-Pro stands out for the solidity of its results when generating images based on text instructions (PROMPT), reflecting precise and consistent details aligned with the description. This level of performance has been achieved largely thanks to improvements in training processes and a careful balance of data between the synthetic and real worlds.
What improvements does Janus Pro 7B introduce?
- Expansion of the dataset
DeepSeek’s technical report indicates that 72 million high-quality synthetic images have been added, carefully balanced with data from the “physical world.” The goal is for the model to acquire a more comprehensive understanding of reality and visual aesthetics, achieving more coherent and appealing results. - Increased model size
The Janus-Pro-7B version refers to the 7 billion parameters that make up its core. This increase in capacity translates to better retention of details and nuances in each generated image. Despite the size, DeepSeek claims to have optimized the model’s training and execution speed to make it viable in production environments. - Improved image stability
One common criticism of AI image generation is instability or inconsistencies in details, especially in complex scenarios. DeepSeek asserts that Janus-Pro has significantly improved in this area, reducing issues such as distortions in human faces or inconsistencies in object dimensions. - Intelligent feedback algorithms
Although not all details have been disclosed, the report mentions the adoption of continuous feedback techniques, enabling the model to learn from its own “errors” and reinforce the most effective patterns for recreating accurate images.
How to access or test Janus-Pro?
DeepSeek upholds an open-source philosophy for most of its products, including the Janus family. This means that, in principle, interested developers can download the models, review the code, and experiment with them in their own environments.
If you want to learn how this AI model works, click on this link.
What is DeepSeek?
DeepSeek is a Chinese Artificial Intelligence startup that has recently caught the industry’s attention thanks to its remarkable advancements.
Under the leadership of Liang Wenfeng —an expert in quantitative investment funds— the company recently launched an Artificial Intelligence model that, according to various analysts, delivers results comparable to those developed by giants like OpenAI and Meta.
This would not be news if it weren’t for the fact that the training cost and infrastructure required for its AI are reportedly much lower than those of its Western competitors.
DeepSeek managed to secure the top spot among the most downloaded free apps on Apple’s App Store, contributing to increased media pressure and curiosity to learn more about the technology behind its AI. Additionally, its open-source philosophy has raised expectations in the community by allowing developers worldwide to analyze and leverage its models.
