
Don't use it to post naughty pictures online please.
The v1.9 version uses a custom LORA block arrange you can copy it if you want.
LCM does work. But it sacrifices overall detail and promp coherence.
Recommendations
This can be used to generate images of 768/768, 768/512, 512/512 and anything in between, in all my testing no duplicate faces were generated.
Best CFG 7 - 9, i would not go any higher unless you wanna pump up the steps too.
Any checkpoint works but you may have to adjust the weight in order to get it to look right, the max usable weight is 1.3.
I recommend Restart/DPM++ 2M for the first pass and Restart/DPM++ 3M SDE for the Hires.fix pass, this will give you the best results.
Don't go above 2.0 times the first pass resolution, is better to pass it to img2img or use some other tricks to obtain more detail.
Extra notes.
The max resolution i tested is 960, it looks good, but expect duplicated subjects and faces.
描述:
Similar dataset, but the worse pictures were replaced.
New training process, all the bell and whistles for SD1.5 except for Debiased Estimation loss.
Further improvements via caption weights.
训练词语: lesydx
名称: LeSydx.safetensors
大小 (KB): 117715
类型: Model
Pickle 扫描结果: Success
Pickle 扫描信息: No Pickle imports
病毒扫描结果: Success