Thank you to razzz for the tip on training to 800mb!

14400 steps. Yep this one was a big one. Trained in local Kohya install. 4090. Trained on DreamShaper XL1.0 Alpha2. I followed SECourses SDXL LoRA Guide.

This model was testing a couple of things. One being the larger training image dataset, hence the larger amount of steps. Two being, yes, a non-1.7GB LoRA. Only future tests/outputs/comparisons will tell if there is a quality loss in this or not. I decided to scale down the size per high demand and preliminary speculation based on observations that hopefully it wont diminish the overall quality of the model. Worth noting if you are also training SDXL LoRAs yourself that this method of training to achieve 800mb files also lowers your overall GPU VRAM usage from ~18gb to ~15gb.

Most images were on DreamShaper XL A2 in A1111/ComfyUI. Hi-res fix with R-ESRGAN (1.25) and 0.2-0.4 denoising strength. Upscaled using "4x_NickelbackFS_72000_G" or "4x_NMKD-Siax_200k"

Help me out by liking and following me to see my new models. Submit some of your outputs and post your SUGGESTIONS in the comments below. I am pacing around 6-10 models weekly depending on my actual job workload. If you liked my models and want to support the work I do you can buy me a pizza!

List of upcoming LoRAs I intend to do as well as what I am currently working on.

Thank you to razzz for the tip on training to 800mb!