3 ms·
> but the community is so entrenched in Stable Diffusion 1.5 and the old Stability AI PyTorch implementation that moving to anything else hasn't really happened
by nullc 3y ago
> but the community is so entrenched in Stable Diffusion 1.5 and the old Stability AI PyTorch implementation that moving to anything else hasn't really happened yet.
It's not just inertia, the later models (even 1.5 over 1.4) are heavily censored. And at least in the case of 1.5 it appears to cause obviously inferior output even for content which is in no way explicit.
My understanding is the facebook published Llama 2 chat fine tunes are more or less the most censored LLM yet, refusing to even discuss negative sentiments. I bet we'll see other people's RLHF fine tunes on the base model being used as a point for more developent.
It's an unfortunate development that it appears that the primary commercial "value add" is building models that don't do what the users unambiguously direct them to do. It's hard to make a model smarter, but easier to lobotomize one.
- brucethemoose2 3y ago> the later models (even 1.5 over 1.4) are heavily censored. And at least in the case of 1.5 it appears to cause obviously inferior output even for content which is in no way explicit. 1.5 is better than 1.4, but 2.0/2.1 are indeed heavily lobotomized. The jury is still out on SDXL, but I think its much better. But the second thing I alluded to, the backend, is going to be more of an issue. The chat finetunes of Llama 2 are kinda irrelevant because the community will come up with their own very uncensored finetunes.
- nullc 3y agoWell automatic1111 does SDXL and the model is clearly very censored-- now perhaps the censorship doesn't handicap its performance as much as it did with 2.0... that's harder to gauge without an uncensored comparison point.