SDXL 1.0 vs SD 1.5 Character Training using LoRA Kohya ss for Stable diffusion Comparison and guide

#stablediffusion #A1111 #AI #Lora #koyass #sd #sdxl #character #charactertraining This video shows and presents the steps needed for a Perfect LoRA Model of a character that is flexible, able to adapt to new settings, and works for face and full body shots with high level of details for both SD 1.5 and SDXL. bellow are the steps and key notes used and some of the resources used in the process. LoRA Training Guide: Training Steps After installation of Koyass from https://github.com/bmaltais/kohya_ss Prepare your data set (Good Data is essential) Caption it Set up folders Set up training parameters Train for few epochs often 5 – 10 epochs could be enough (or more) select best epoch after comparison. Test for each epoch: new set of clothes, different colors Conclusion: - SD 1.5 can produce hyper realistic images with 1024x1024 image same as SDXL - after detailer is required in most full body shots when having lower quality full body image training data - using lower number of images (around 20) can work well with high resolution data set - SDXL 1.0 produces finer more detailed skin features and slighly more realistic subjects - SDXL 1.0: it is recommended to use smaller network ranks, such as 16 or 8 to reduce file sizes - SDXL file size = 6 times SD 1.5 for same network rank - Regularization set: use High quality set with realistic images and of same dimensionj as training data - SDXL: better use 768x1344 images to improve quality of full body shots - SDXL: use Gradient checking for lower VRAM - SDXL training parameters and methodoly is almost same as 1.5 except with minor difference (network rank, SDXL check point) - Biggest obstacle for SDXL is: Higher VRAM requirements. other useful info about how stable diffusion works in general and some tips can be seen at https://youtu.be/1MkZB2YEiy4 Beginners guide to stable diffusion in Automatic1111 at: https://youtu.be/RtjDswbSEEY cloth/object training https://youtu.be/wJX4bBtDr9Y Koyass for lora training https://github.com/bmaltais/kohya_ss Regularization set: it is recommended to have high quality reg set of the same class, same dimension. you can download these models from https://civitai.com/models/126328/katheryn-winnick-xyzkwv1sdxl you can download a sample 460 regularization images with 1024x1024 from this link which are mostly generated from SD and resized using Topaz AI software and some from freepik free images https://huggingface.co/datasets/AIHowto/460RegImages1024x1024/resolve/main/reg_images_1024x1024jpg.zip For local tests: Adapter Type NVIDIA GeForce RTX 3070 Laptop GPU, NVIDIA compatible, 8GB VRAM Physical Memory (RAM) 16.0 GB Processor AMD Ryzen 7 5800H 3201 Mhz, 8 Core(s), 16 Logical Processor(s)