Original description: *THIS CHECKPOINT IS SDXL1.0 About this version: Initial release, base on CounterFeit v3 ----- How-to a. Positive Prompts: - Like BArtstyle LoRA, color scheme still work, and I managed to make these typical color like blue theme or pink theme much closer to original data without overfit. - Natural Language + Low CFG (4.5-6) + Roll dices + cherry-pick > "quality tags" - From my test, about 30% image should be good enough first glance. If you can't get similar result, check if there are any duplicate/conflict tags in your prompt, shrink it + lower your CFG. b. Neg Prompts: - DO NOT USE FCKING LONG NEG PROMPTS UNLESS YOU KNOW WHAT ARE YOU DOING! - For random generate, simple EasyNegativeV2 is good enough. - Try to add 'glowing, 3d, painting \(medium\), photorealistic, nsfw' in negative prompts if you think this model is still too much CounterFeit style, but can't guarantee the effect. c. VAE: WD1-4-kl-f8-anime2-bless09.ckpt d. Sampler: Euler A - step 50-100 - CHAD Euler A still better than Virgin DPM e. Upscaler: AnimeVideo v3 - step 10-15 - Denoising 0.45-0.55 ----- Base Model: a. Counterfeit V3-fp16-fix - After 6 MONTHS, author has released a vae fix version of CFv3. OG version had a serious VAE issue that make all the LoRA of CFv3 is bit broken, even if you replace the VAE won't help. Recommend to update if you use it. - It seems like it is the ONLY TRAINING model that is capable with NAI 1.0 LoRA. That means only this model is good for fine-tuning b. Dataset: 200 images from Blue Archive's Official Artwork + flip, 400 in total. - TOTALLY OVERHAULED DATASET COMPARE WITH BArtstyle LoRA. - Highest resolution I could found, cropped to improve dataset quality - Use XnConvert Downscale to 1280x1280 c. Training Setting: - OPTIMIZER: AdamW8bit - SCHEDULER: Constant with Warmup - Warmup to 10% of total steps. - UNET LR: 1e-6 - TE LR: 1.25-7 - 1/8 of UNET LR as suggestion - Stop TE Training at 50% of total steps. - EPOCH: 100 - BATCH SIZE: 1 - STEP: 40000 - RES: 1280X1280 - TIME: 10H27M