experimental model of kokorone awayuki (心音淡雪) from 「VTuberなんだが配信切り忘れたら伝説になってた」
[Update log] :
[2024-07-12] : release v2.0 -> full model retrain at higher resolution using data from. model took longer to train than expected, figuring out how to optimize.
this version took ~5 hours to train, 2 epochs @ 2500 steps (5000 total).the sample images for v2.0 were generated with a VAE. it looks more appealing to those browsing models (i forgot to use a VAE).
I have a PonyDiffusion/SDXL derivative LoRA trained, but have no idea how positive and negative prompting work to get my preference of style to carry over. I have no plans to completely migrate to SDXL.
[2024-07-09] : release v1.5 -> refresh of first model retrained with additional data, fixed activation tags.
this version took ~30 minutes to train, 3 epochs @ 2500 steps (7500 total). this version may also have been overtrained as it continued training using the previous model as a weight; generations still look decent.[2024-07-08] : release 1.0 -> initial version of model, retrain later to fix activation tags and refine detail.
this version took ~40 minutes to train, 3 epochs @ 1600 steps (4800 total).
[Recommended LoRA strength] :
0.7 or 0.8[Recommended prompt] :
1girl, <lora:kokorone_awayuki:0.8>, kokorone_awayuki, black hair, white hair strand, pink eyes, long white dress, long dress, white elbow gloves, flower, white gloves, snowflake earrings, hair flower, hair ornament, long hair, off-shoulder dress, off shoulder, solo, tiara, virtual youtuber, fur trim
thank you for browsing, downloading, and generating!