Hm, they still have work to do on that front. For comparison, left is deepseek, right is dall-e (you can run DS’ Janus7B on huggingface, and dall-e I ran from chatgpt directly):
I gave both the same prompt (which was suggested on huggingface) but after the first generation from dall-e I told it to “make the astronaut face the camera” so I could get similar end products.
The interesting part with deepseek’s R1 is that it’s so much more efficient than GPT4o. I didn’t look far enough to see how Janus compares but that would be the real nail in the coffin – image generation at a fiftieth of the energy expenditure. Image generation so cheap you don’t need to pay for it.
Mind you I hate image gen AI lol, but overall it’s beneficial to destroy the western tech industry.
Lol bravo!
also be sure to join the DeepSeek sub https://lemmygrad.ml/c/deepseek