These are beta versions. They have issues, so be prepared to work for it. I didn't go through the keywords at all.In a future version I guess I need to train each thing different chair in Kohya_ss? Ones with hands and ones without perhaps? Different folders of training data perhaps in same Lora? Not sure.I still don't know what I'm doing.These were trained on "CyberRealistic v4.2".I was using a weight of 1. I rarely even checked other weights.---------------------------------------v0.2 - Beta (more consistent results, less variety)Trigger Word: ch41r_91rlTraining Images: 36 imagesI went with epoch 4. I couldn't decide what worked best.1,440 steps for 10 epochs.576 steps for 4 epochs.The training data contains all 10 epochs. Epoch 3 was also good.This focuses on arm chairs. You can try to force a plastic or wood chair, but other chairs are unlikely.I edited a lot more pictures manually in this version. Some of that means the imagery will sometimes be distorted. I'll have to remove some of those images and add others eventually.---------------------------------------v0.1 - Beta (more variety, but more wacky)Trigger Word: ch41r_91rlTraining Images: 173 imagesI went with epoch 3. I couldn't decide what worked best.3,460 steps for 10 epochs.1,038 steps for 3 epochs.The training data contains all 10 epochs. This version include a lot of images from Image Creator from Microsoft Designer (formerly Bing Image Creator). It also contains a lot of hands on the arm rests. It also contains some office chairs. There were a lot of images that were not as good in this one.There are a lot of plastic chairs. Wood was probably second. And then armchairs, which I really like best but didn't have a lot of, then office and leather chairs as well. The images from Microsoft's Image Creator have a lot of flat images of women on the chairs, with hands included. Basically pictures in some cases.I added all the different types of chairs under one tag. I still don't know what I am doing. So, it randomly will give you a kind of chair girl. It's not great. I think some of the lower ones occasionally had good images too.

Files with SHA256: d06de2fc2892ae100982f97a8f2ed8822f0ff4d49c19a9c0d344c231f1494d9e