Search Results
7/16/2025, 3:08:33 AM
>>8903179
SDXL is also the basis of BigASP, Lustify and their hybrid BigLust, which are currently the most popular checkpoints for realistic porn. A group in Korea trained their own SDXL-based anime checkpoint called Illustrious which doesn't obfuscate artists. There was a further refinement of Illustrious called NoobAI, and derivatives of NoobAI are now the most common checkpoints used on /h/.
In the meantime, Stability AI started being opposed to having their models used to generate porn, so they released Stable Diffusion 2 and 3, which were practically useless: SD3 was famously unable to generate a woman lying on grass without messing up the anatomy. Some employees of Stability broke away and founded a new company in Germany called Black Forest Labs, which released Flux, the current state-of-the-art local model. It doesn't actively censor porn, but it still didn't have much NSFW in its dataset, so it needs more training. The most promising of these efforts is Chroma, which is trying to finetune Flux with anime, furry, photo and classical art datasets. This is the model that was used for all the impressionistic gens in the last thread. Chroma doesn't seem to be able to identify anime artist styles by name, so /h/ is waiting to see if that'll improve when its training is supposed to be finished at epoch 50.
Also, the video generation models that came out of China -- Wan and Hailuo -- can be used to generate images if you just set them to produce one frame.
SDXL is also the basis of BigASP, Lustify and their hybrid BigLust, which are currently the most popular checkpoints for realistic porn. A group in Korea trained their own SDXL-based anime checkpoint called Illustrious which doesn't obfuscate artists. There was a further refinement of Illustrious called NoobAI, and derivatives of NoobAI are now the most common checkpoints used on /h/.
In the meantime, Stability AI started being opposed to having their models used to generate porn, so they released Stable Diffusion 2 and 3, which were practically useless: SD3 was famously unable to generate a woman lying on grass without messing up the anatomy. Some employees of Stability broke away and founded a new company in Germany called Black Forest Labs, which released Flux, the current state-of-the-art local model. It doesn't actively censor porn, but it still didn't have much NSFW in its dataset, so it needs more training. The most promising of these efforts is Chroma, which is trying to finetune Flux with anime, furry, photo and classical art datasets. This is the model that was used for all the impressionistic gens in the last thread. Chroma doesn't seem to be able to identify anime artist styles by name, so /h/ is waiting to see if that'll improve when its training is supposed to be finished at epoch 50.
Also, the video generation models that came out of China -- Wan and Hailuo -- can be used to generate images if you just set them to produce one frame.
Page 1