Beta136: Improved dataset and reduced Learning Rate==========================================Beta107: Replace the refined data set from Pony==========================================Warning! Beta130, Beta122, and Beta92 v1.0 was created as part of the research on AdamW 32 bits, and not intended for regular usage.Please add the following arguments in Automatic1111 WebUI webui-user.bat in order to enable 32 bits.<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/17961e5c-527e-4f5d-b1ce-216324894022/width=525/17961e5c-527e-4f5d-b1ce-216324894022.jpeg" />Feel free to create the picture if interested, however, the file is huge.Thanks to the members from the following groups for generating the pictures and discussing the result, so that this LoRA could be improved greatly before released:AI乐园: <a target="_blank" rel="ugc" href="https://t.me/+1XT8FjX5CUlhNDNl">https://t.me/+1XT8FjX5CUlhNDNl</a>AI绘图魔法群: <a target="_blank" rel="ugc" href="https://t.me/+jPp0_tAk-UY5NDhh">https://t.me/+jPp0_tAk-UY5NDhh</a>AI练丹交流奇思妙想会AI绘画lora交流AIGenerateMagic兔狲的SDXL百老汇============================================Thanks for the interest in this LoRA creation. This project has been conducted in Stable Diffusion 1.5 since March 2023.The objective of this project is to create an ideal training data set for LoRA. After the launching of SDXL, the scope of this research was expanded to this new version.The face of the researcher was taken as reference in order to create the training set. 15 photos were taken purposely for AI training only. These 15 photos contain different angles of face and shoulder to fulfill the AI training requirement.After trying different combinations of settings, an optimal adjustment was found.<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/2cb302ad-5a2f-4f00-93b8-4cbea34ca5fa/width=525/2cb302ad-5a2f-4f00-93b8-4cbea34ca5fa.jpeg" />Comparison of Dim256, Dim128, and no LoRA respectively.*********************************************************Here is the log for different version.<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/ad76e8c7-d17f-4d5b-8672-ecc28d479da1/width=525/ad76e8c7-d17f-4d5b-8672-ecc28d479da1.jpeg" />Beta1: First trying with suggested learning rate 4e-7, resolution 1024x1024, dimension 256, alpha 256, optimizer: Adafactor. Result unsatisfied.Beta2: Change the learning rate in Beta1 to 0.0004. Result unsatisfied.Beta3: Reduce the learning rate in Beta2 to 0.0001, increase the steps of each photo to 100. Result unsatisfied.Beta4: Replace the resolution in Beta3 to 2048x2048, reduce the steps of each photo to 20. Result unsatisfied.Beta5: Increase the steps of each photo to 100 in Beta4. Result unsatisfied.Beta6: Applying learning rate 0.0004, resolution 1024x1024, 20 steps of each photo, alpha 1. Result unsatisfied.Beta7: Reduce learning rate to 0.0001, resolution 960x960, alpha 256. Result unsatisfied.Beta8: Reduce learning rate to 0.00001, resolution 1024x1024, 20 steps of each photo. Result unsatisfied.Beta9: Reduce learning rate to 0.000005. Result unsatisfied.Beta10: Reduce resolution 512x512, 40 steps of each photo. Result unsatisfiedBeta11: Using two subset of resolutions 512x512 and 1024x1024, 40 steps of each photo. Set max resolution to 512x512, dimension 128, alpha 128. Result unsatisfied.Beta12: Set Max resolution to 1024x1024, dimension 64, alpha 64. Result unsatisfied.Beta13: Replace optimizer to prodigy, learning rate 1, dimension 128, alpha 128. Result unsatisfiedBeta14: Replace optimizer to adamW8bits, learning rate 0.0001, 20 steps, dimension 256, alpha 256. Result satisfied.Beta15: Reduce the learning rate to 0.000005. Result satisfied. Uploaded to Civitai as version 1. Marked as miletone.Beta16: Increase the learning rate to 0.00001, no better results.Beta17: Reduce the learning rate to 0.000005, dimension 128, alpha 128. Applying Text Encoder learning rate to 0.000005, Unet learning rate to 0.00001. No better results.Beta18: Removed Text Encoder learning rate and Unet learning rate. No better results.Beta19: Increased learning rate to 0.0001, resolution 512x512 only, 80 steps, dimension 256, alpha 256, clip skip 2. No better results.Beta20: Reduced learning rate to 0.00001. No better results.Beta21: Increased resolution to 960x960. No better results.Beta22: Reduced learning rate to 0.000001, resolution 1024x1024, 100 steps. No better results.Beta23: Reduced learning rate to 0.0000004, 400 steps. No better results.Beta24: Reduced steps to 200, dimension 64, alpha 64, skip clip 1. No better results.Beta25: Increased learning rate to 0.00001, 30 steps, dimension 256, alpha 256. Result satisfying. Marked as second milestone.Beta26: Reduced learning rate to 0.000005, 20 steps, No better results.Beta27: Increased learning rate to 0.00001, 30 steps, dimension 64, alpha 64. Result satisfying.Beta28: Replace new training data set. Uploaded to Civitai as Dim64 version.Beta29: Increased dimension to 256, alpha 256. No better results.Beta30: Increased resolution to 1536x1536. No better resultsBeta31: Increased resolution to 2048x2048. No better resultsBeta32: Changed resolution to 1536x2048. No better resultsBeta33: Reduced resolution to 1280x1280, 40 steps. No better resultsBeta34: Changed resolution to 1024x1536, 30 steps, new set of the training data. No better results.Beta35: Changed resolution to 1024x1280. No better results.Beta36: Changed resolution to 1024x1024. No better results.Beta37: Changed resolution to 1024x1024, 50 steps. No better results.*********************************************************Since it is ongoing project, your valuable feedback is important, in order to bring this project to the next level.

kwFemale SDXL

The Kohya_ss settings are as follows:Optimizer: AdamWMix Precision: noSave Precision: floatLearning rate: 5e-6Unet LR: 5e-6Text Encoder: 2.5e-6Dimension: 256Alpha: 256Max Resolution: 1024,1536

Files with SHA256: c78a7343d2a6763d5ebb4cd97d90d6aebf8371d1b483806de49a03f012df385e