Type | |
Stats | 1,267 |
Reviews | (241) |
Published | Oct 2, 2024 |
Base Model | |
Training | Steps: 4,636,800 Epochs: 40 |
Usage Tips | Clip Skip: 1 |
Hash | AutoV2 54C8530E02 |
Hi there and welcome to another checkpoint using my finetune based on IllustriousXLv0.1, aimed at having 500+ working artist styles reinforcing their knowledge and reducing the Lanczos / Nearest resize artifacting shown by base Illustrious, this was meant as a test vehicle for reducing artifacting across all artists and not dilute the knowledge of already trained ones, as such work cleaning signatures/logos/urls wasn't done, if you consider this finetune has potential and would like to see an even bigger one, please consider donating to my Ko-fi,
Please use the style triggers to change how the checkpoint looks, in some samples CFG can go up to 8.
Two versions: v1 and LoRA Extract
Artist tokens work best before the score tags, but you can safely use them at the end
You can find the list of artists here(Some might not work, as Clip L for Illustrious is quite toasty sometimes it NaNs on float precision lmao), style mixing works just have to play with weights, also you can put them in negatives to further guide how you want it to look. Emphasis mode No Norm is suggested for SDXL models
Huge thanks to @Anzhc with his help in training hyperparameters, to @FallenIncursio , @mfcg, @mnemosynekotorin , @richyrich515 , @justTNP, and @Kojimbomber they were kind enough to provide much needed data for this!
Positive prompt : style, masterpiece, best quality,
Negative prompt : worst quality, low quality, jpeg artifacts, multiple views, 4koma, bad anatomy, bad hands, text
Personal schizo Negative prompt: lowres, worst quality, low quality, bad anatomy, bad hands, multiple views, 4koma, comic, greyscale, censored, jpeg artifacts, patreon logo, patreon username, web address, signature, watermark, text, logo, artist name, overly saturared, overly vivid
Sampler : Euler A AYS, Any Align Your Steps, SGM Uniform, Euler A CFG++
CFG : 4 to 7 (can be pushed up to 8 and it will work on a lot of prompts
If you want to support me, consider liking the model, and/or tipping buzz/donating to my Ko-fi, as this was finetuned on 8xA6000 Ada (Rought stimate is 400 USD for this besides tagging,and testing locally before training)
Relevant training information:
Trained with at least 130k images.
Total of 4.636.800 real steps after batching across 30 epochs.
Effective batch size was 22 x 8 GPUs x 2 Gradient Accumulation
Learning rate was 2.5e-6, TE1 was Frozen