LoRA vs LyCORIS and Regularization comparisons in Kohya ss | stable diffusion person training Part 2 скачать в хорошем качестве

LoRA vs LyCORIS and Regularization comparisons in Kohya ss | stable diffusion person training Part 2

This video is Part 2 of LoRA trainging for stable diffusion, it focuses mainly on comparisons between LoRA / LyCORIS using regularization set and without First: sorry for the bad organization of this video, did my best within the time available. the video will have the following content: How LyCORIS training is performed Comparison (standard LoRA vs LyCORIS) for same dataset Comparing Results using regularization sets and without How many reg images to use in Kohya ss? Results with captioned subjects and without captions Conclusion (based on the assessed models) LyCORIS/LoCon gave slightly better results for Person training than Standard LoRA Reducing Alpha is useful to make training finer and reduce overfitting and produced better results. For small data sets of people, such as under 40 images, having 128 Dim is not good, it seems that (Dim=64, Alpha=32 and smaller values for Standard LoRA) or (Dim=32x alpha=8 -- convolution 4x convolution alpha=1 for LyCORIS) gave better results than higher alpha and network dimensions Regularization results were not conclusive , better in some cases but not in others, it makes the model slightly more flexible for new clothes/poses/eye color/hair color, body types, but in most portraits, having no regularization seems to give better resemblance… so ideally, if we have time we should develop two models with reg and without and see which is better…if you have little time, don’t use regularization because its faster to train Number of Regularization images : Kohya ss seems to only uses (image repeats * image count ) class images and ignores the rest Caption is better than not captioning, without captioning, the image becomes less flexible and have less resemblance with new clothes/settings sample regularization images I used before sample 1832 https://huggingface.co/datasets/AIHow... sub sample 1340 used in some tests too https://huggingface.co/datasets/AIHow... any data set could be used as reg images as long as it resembles the subjects with different variations, one can use images generated by SD or real images too, ideally based on many notes from others, generated from same SD... with general colors/saturation levels...etc. To generate your Reg images: from you A1111 SD: choose your target checkpoint, put a woman (or your class name) and generate few hundreds of images ... if your data has full body shots, also generate few hundreds full body shots, you can use a suitable commands such as (realistic woman, master piece, best quality, young woman , full body )... and few hundreds upper body shots...so that it is mixed and resembles your training data in terms of variety... Kohya will eventually only use (repeats*image count) from your class folder and ignores the rest ... this applies for any other class. Kohya also includes another method for regularizatio besides using images which is Scale weight norms in parameter's tab, didnt test it yet though. #stablediffusion #a1111 #kohyass #kohya #lora #lycoris #modeltraining #ai #fashion