News for hackers. Stuff that matters.
Re:DeepSeek releases Janus Pro, a text-to-image generator [pdf] by m348e912 on Monday January 27, 2025 @02:13PM (#42844540)
I'm no AI model expert, but it looks like a number of DeepSeek models have been modified to remove the censorship restrictions and uploaded to huggingface. Perhaps we will see an uncensored version of Janus Pro soon.

https://huggingface.co/models?sort=created&search=deepseek+u...

Re:DeepSeek releases Janus Pro, a text-to-image generator [pdf] by whimsicalism on Monday January 27, 2025 @03:02PM (#42845012)
those will come, but it's worth noting none of these models are actually r1-derived, they are based on the distills
Re:DeepSeek releases Janus Pro, a text-to-image generator [pdf] by bilbo0s on Monday January 27, 2025 @02:26PM (#42844662)
No.

The Chinese just provide models aligned to global standards for use outside China. (Note, I didn't say the provided models were uncensored. Just that it wouldn't have so much of the Chinese censorship. Obviously, the male-female question in the original comment demonstrates clearly that there is still alignment going on. It's just that the alignment is alignment to maybe western censorship standards.) There is no need to modify DeepSeek at all if you want non-Chinese alignment.

Re:DeepSeek releases Janus Pro, a text-to-image generator [pdf] by zb3 on Monday January 27, 2025 @02:41PM (#42844823)
Actually I wish these models had Chinese alignment without the western one..
Re:DeepSeek releases Janus Pro, a text-to-image generator [pdf] by bilbo0s on Monday January 27, 2025 @02:47PM (#42844887)
Um, yeah.

Pretty sure that's not gonna be an option for you. At least not in the US.

Re:DeepSeek releases Janus Pro, a text-to-image generator [pdf] by maxloh on Tuesday January 28, 2025 @03:11AM (#42850001)
Any idea how was those censorships removed?