Re:DeepSeek releases Janus Pro, a text-to-image generator [pdf]
by noja on Monday January 27, 2025 @01:41PM
(#42844228)
All models do this last time I checked. Not just Chinese.
[ Parent
| on HN ]
Re:DeepSeek releases Janus Pro, a text-to-image generator [pdf]
by w4yai on Monday January 27, 2025 @01:47PM
(#42844290)
Wrong, at least with Claude : https://i.imgur.com/6dj1XAU.png
Re:DeepSeek releases Janus Pro, a text-to-image generator [pdf]
by Palmik on Monday January 27, 2025 @01:57PM
(#42844400)
It's actually not wrong, here is the example from the OP: https://imgur.com/a/5uMYI44
Re:DeepSeek releases Janus Pro, a text-to-image generator [pdf]
by rsanek on Monday January 27, 2025 @07:50PM
(#42847656)
You get the same output from Claude regardless of gender. I wonder which government asked Anthropic to censor the output in this way
Re:DeepSeek releases Janus Pro, a text-to-image generator [pdf]
by girvo on Monday January 27, 2025 @08:40PM
(#42847977)
What government demanded them to do that under pain of being disappeared?
Re:DeepSeek releases Janus Pro, a text-to-image generator [pdf]
by themgt on Monday January 27, 2025 @02:29PM
(#42844692)
"That's nothing. I can stand in front of the Kremlin and yell, 'To hell with Ronald Reagan,' too."
Re:DeepSeek releases Janus Pro, a text-to-image generator [pdf]
by hdjjhhvvhga on Monday January 27, 2025 @02:27PM
(#42844675)
Yes, but one would expect the Chinese not to fine-tune according to Western standards like in the quoted prompt.
Re:DeepSeek releases Janus Pro, a text-to-image generator [pdf]
by girvo on Monday January 27, 2025 @08:41PM
(#42847990)
It becomes clearer when one realises that it was RLHF'd using ChatGPT and Qwen (and others I assume) output. It's caused it to "learn" the same weird Western censoring, and adopt the China-styled censoring too.