Search Results
7/16/2025, 2:49:50 AM
>>8902214
>SAI? as in paint tool SAI or whatever fro like a decade ago? they made this? thats crazy. im surprised they released this for free and never charged people for it. What was in it for them?
No, not Paint Tool SAI, we're talking about Stability AI, the company that kicked off the whole AI imagegen revolution in 2022 when they released the Stable Diffusion 1.4 checkpoint. 1.4 was a very primitive model and imagegen didn't start being taken seriously until 1.5. There's a company called NovelAI that runs a paid subscription service to generate anime images, and they produced their own checkpoint, NovelAI 1.0, by training with the contents of the Danbooru anime imageboard on top of SD 1.5. And this checkpoint got leaked to the public, which was a big deal for /h/. Notable checkpoints that were further refinements of NovelAI 1 include Based64 through 68.
SD 1.4 and 1.5 were trained to generate 512x768 images, but the quality could be improved through upscaling or HiRes Fix (which this is an example of). Stability AI then introduced Stable Diffusion XL, which could generate 1024x1024 (but in practice, most people use 1216x832 and continue to upscale from there). SDXL became the basis of Pony Diffusion XL by AstraliteHeart, which includes datasets from Danbooru, Derpibooru (the pony imageboards) and Rule34. There's a big ecosystem surrounding Pony, but it's controversial because AstraliteHeart hashed the names of all the artists in the training data, ostensibly because he didn't want users to replicate an artist's style exactly, but also because he runs a subscription service and he wants to charge money to unlock the names.
>SAI? as in paint tool SAI or whatever fro like a decade ago? they made this? thats crazy. im surprised they released this for free and never charged people for it. What was in it for them?
No, not Paint Tool SAI, we're talking about Stability AI, the company that kicked off the whole AI imagegen revolution in 2022 when they released the Stable Diffusion 1.4 checkpoint. 1.4 was a very primitive model and imagegen didn't start being taken seriously until 1.5. There's a company called NovelAI that runs a paid subscription service to generate anime images, and they produced their own checkpoint, NovelAI 1.0, by training with the contents of the Danbooru anime imageboard on top of SD 1.5. And this checkpoint got leaked to the public, which was a big deal for /h/. Notable checkpoints that were further refinements of NovelAI 1 include Based64 through 68.
SD 1.4 and 1.5 were trained to generate 512x768 images, but the quality could be improved through upscaling or HiRes Fix (which this is an example of). Stability AI then introduced Stable Diffusion XL, which could generate 1024x1024 (but in practice, most people use 1216x832 and continue to upscale from there). SDXL became the basis of Pony Diffusion XL by AstraliteHeart, which includes datasets from Danbooru, Derpibooru (the pony imageboards) and Rule34. There's a big ecosystem surrounding Pony, but it's controversial because AstraliteHeart hashed the names of all the artists in the training data, ostensibly because he didn't want users to replicate an artist's style exactly, but also because he runs a subscription service and he wants to charge money to unlock the names.
Page 1