5 ms·
For practical context, Stable Diffusion 2.X was trained on LAION-5B as opposed to LAION-400M for Stable Diffusion 1.X. At the least, Stable Diffusion 2.X is be
by minimaxir 4y ago
For practical context, Stable Diffusion 2.X was trained on LAION-5B as opposed to LAION-400M for Stable Diffusion 1.X.
At the least, Stable Diffusion 2.X is better at pain points of image generation such as text legibility and hands, potentially due to having more data points.
- Terretta 4y agoProblem with hands is probability. It's more probable a finger has a finger on both sides of it than not. So the model diffuses lots of adjacent fingers.
- stavros 4y agoBut that's the same for everything that has structure. A small section of an arm is much more likely to have another small section of an arm next to it than to have a hand, yet SD's arms are usually well-proportioned.
- sdenton4 4y agoThere's a lot of loooooong necks, though.
- alar44 4y agoNo the problem with fingers is that they resemble hotdogs and the AI really likes hotdogs so you get a lot of fingers. I can make things up too!
- satvikpendem 4y agoSD 2 also removed quite a lot of images of humans due to their fear of people generating CSAM, so the quality actually has gotten worse for anything resembling humans than SD 1.
- astrange 4y ago2.0 removed too many of them due to a bug in the NSFW filter. 2.1+ should be better again. But they’re harder to control without negative prompting.
- in3d 4y agoThis is incorrect. Stable Diffusion 1.x was trained on "laion-improved-aesthetics" (a subset of laion2B-en).
- minimaxir 4y agoDouble checked and both the initial comment and the correction are incorrect: the original v1.1 was trained on LAION-2B, then subsequent versions were finetuned on the aestethics subset. Either way, the main point is the same: more training data gives better results. https://github.com/CompVis/stable-diffusion#weights https://github.com/CompVis/stable-diffusion#weights