When you purchase through links on our site, we may earn an affiliate commission. Here’s how it works.

DeepSeek's Janus-Pro model outperforms OpenAI's DALL-E 3 in image generation

DeepSeek
image by Mojahid Mottakin via Depositphotos

DeepSeek has made waves on the internet after reports claiming that it outperforms most advanced AI models worldwide started circulating on the internet. Remarkably, DeepSeek spent less than $6 million to train its AI models, compared to $100 million invested by OpenAI for the same. This resulted in not only DeepSeek becoming the number one free app on the App Store but also triggered a historic wipeout of over $400 billion in NVIDIA's market cap in the US. The AI chatbot saw overwhelming traffic causing DeepSeek server outages and performance problems and the firm has also blamed cyberattack.

Despite these challenges, the Chinese AI lab has forged ahead, unveiling a groundbreaking open-source AI model called Janus-Pro. The new model is already making news, because according to reports, the Janus-Pro image generation model, outperforms OpenAI's DALL-E, Stability AI's Stable Diffusion, and other image generation models in several benchmarks.

Janus-Pro is an update over Janus, which was launched late last year. Janus-Pro comes in various sizes, from compact 1 billion parameters to the 7 billion parameters version, near the size of SD 3.5L. The largest model, Janus-Pro-7B, excels in both image generation and analysis, beating top competitors PixArt-alpha, Emu3-Gen, and SDXL on industry benchmarks GenEval and DPG-Bench, according to information shared by DeepSeek. The Janus-Pro-7B model is available to download for free from Huggingface, a popular place for AI and machine learning.

Janus-Pro-7B DeepSeek

Janus-Pro-7B is based on an autoregressive framework that separates visual encoding processes while keeping a unified transformer architecture for processing. This "not only alleviates the conflict between the visual encoder’s roles in understanding and generation but also enhances the framework’s flexibility." However, while Janus-Pro outperforms its competition across multiple tasks, it doesn't outshine specialized models built for singular processes.

Janus-Pro-7B DeepSeek

This new image generation model comes hot on the heels of DeepSeek's earlier success with the R1 language model that is giving tough competition to GPT-4's abilities, for a fraction of money. The low development cost of these advanced models has sent shockwaves in the US AI industry.

Report a problem with article
jbl bar 91 soundbar with subwoofer
Next Article

JBL Bar 9.1 truly wireless Dolby Atmos soundbar with subwoofer is back to lowest price

TikTok Microsoft USA
Previous Article

Microsoft in talks to buy TikTok's US operations, says President Trump

Join the conversation!

Login or Sign Up to read and post a comment.

2 Comments - Add comment