DeepSeek is on a roll. Not content with exploding the apple cart with its ChatGPT-rivaling R1 model, it's just released a new multi-modal model upgrade called Janus Pro.
These new 1B and 7B models can complete image generations and also understand visuals, which is becoming an increasingly important part of modern day AI.
I took a look myself at this latest offering from what's easily the hottest AI company in the world right now.
If you're curious to try it for yourself, you can access the model at HuggingFace here.
The promise
This is the second generation of the Janus model, and it’s supposed to deliver improved image quality, and an ability to handle text.
Another key difference is the fact that the new model combines visual understanding alongside image generation — so it can "see" an uploaded image and understand it.
This is not a typical combination with conventional models. They call it unified multimodal.
The reality (for now)
Unfortunately all this tech seems to have gotten in the way of creating a knockout product.
It’s not that the model is bad so much, it’s just that the image generation feels two years old. Forget about creating human faces; they’re distorted, twisted, and the very worst of early AI image generation. Think about what Stable Diffusion was like in 2023 and you'll know what I'm talking about.
It’s as though we’ve all been whisked back in a time machine to the era of three fingered humans, only it’s now the whole body.
It’s a shame, but I guess innovation often comes with a price. I spent quite a while trying to generate an image which was anywhere near the current state of the art, and failed miserably. You can see the examples below.
The good news is the image vision seems to work fine. I uploaded a shot of someone looking at their mobile phone in a café, and the model accurately depicted what was in the image.
But this is hardly ground-breaking stuff, just about any vision model, proprietary or open source, can do this at the moment. Even the lowly Llava model, which is small enough to run on a home computer, can do this.
Bottom line
So where does that leave us? It’s clear the Chinese have once again tried to innovate with their model design, and on the face of it in a good way. Combining image generation with the ability to read images is a nice feature.
However, the report card on this attempt must read "could try harder."
I’m not sure how or where DeepSeek got the demo images from on its website, and I’m absolutely baffled by the text images the company is boasting about.
Of course these are only tiny models at 1B and 7B parameters, but even so one would hope there would be better output. I got nowhere near the demo results on their site, despite trying different configurations, long prompts and short prompts. It’s a total mystery. I suggest they maybe take a trip back to the drawing board?
More from Tom's Guide
- What is DeepSeek? — everything to know
- Is DeepSeek a national security threat? I asked ChatGPT, Gemini, Perplexity and DeepSeek itself
- I tested the Plaud Note AI voice recorder — here’s my verdict
