Type | |
Stats | 267 0 |
Reviews | (18) |
Published | Mar 21, 2024 |
Base Model | |
Training | Steps: 20,000 Epochs: 20 |
Usage Tips | Clip Skip: 2 |
Trigger Words | all fours seiza wariza squatting kneeling high kick flying kick Monet real life anime +6 more |
Hash | AutoV2 97AABC70C2 |
Spent a literal week training this model locally, since I had great difficulty training it online.
This was trained on 500+ images in 4 specific categories: full body, close up, Victorian, and Monet (the painter, not me lol). Therefore, this checkpoint technically has 4 triggers, just like a LoRA.
This model has the SDXL LCM LoRA and SDXL VAE baked into it, therefore there's no need to use external VAEs or LCM LoRA to get a good result. I'd intended to prune this model, but I'm not sure if it actually did prune it. (To be fair, I'm not sure what a pruned SDXL checkpoint would look like; I simply tried it to see if I couldn't reduce the file size.)
If you're doing a close-up portrait of a face, or simply using ADetailer to refine faces, it's highly recommended to use the "close up" trigger prompt, though it can perform without the trigger word.
I'd had rough experiences with SDXL in the past whenever I tried making images with full body shots, thus I tried to rectify it here. The prompt "full body" is a trigger word which should facilitate the full body landscape shots I'm fond of.
Recommended Usage:
Set the sampler to LCM, reduce the CFG to 1-2, reduce the steps to 8-16, and then generate at will.
If you insist on using non-LCM samplers, then be sure to increase the steps to about 30. But it is of my opinion that this model has worse artifacts than the SD1.5 version when using non-LCM samplers, so be aware of that.