AI & Generative AI
How Good is Deepseek Janus Pro AI - We Test the AI Tool for Images
Much has been written about Deepseek-R1, but not as much about Deepseek's AI image generator Janus-Pro-7B. Kamera & Bild has tested it.
No one has probably missed Deepseek's release of its ChatGPT-like software Deepseek-R1 - which can be used to create, for example, text or program code with generative AI.
Janus Pro 7B, which is their generative AI for images, has been written about less. The program uses an optimization strategy for the images and has been trained on a large amount of data - which we do not know - while the model has been scaled up. Through these steps, Janus Pro is supposed to understand the reasoning that the user wants regarding promptography, that is, creating images through a text prompt.
We have tested it - and it certainly handles certain situations well. People look classically synthetic just like in other models we have tested through generative AI, while the number of fingers and placements end up a bit here and there.
The context doesn't seem to fully work for the prompts because Janus Pro, for example, places a lens in front of a photographer's forehead, something that seems to come from images of photographers who have their camera at their eye. The same goes for images where you see a street where the car drives out from a wall, rather than a road.
Here you can read more about when we tested Adobe Firefly Image 2, or how the development looks for creating AI-generated video sequences with, among others, OpenAI's Sora.
Want to read more about AI? Kamera & Bild has written several articles about AI and generative AI. Among other things, what the law says regarding copyright and AI-generated images.