[ 3 / biz / cgl / ck / diy / fa / ic / jp / lit / sci / vr / vt ] [ index / top / reports ] [ become a patron ] [ status ]
2023-11: Warosu is now out of extended maintenance.

/vt/ - Virtual Youtubers

Search:


View post   

>> No.45072690 [View]
File: 1.22 MB, 2688x3072, xyz_grid-0973-1171692838-best quality, detailed, ouro kronii, ribbon, indoors, dress, hands on own chest,.jpg [View same] [iqdb] [saucenao] [google]
45072690

About HLL4.
I said I'll post it soon, but instead I was experimenting with small test models. I haven't started retraining it yet.
I'll post a new model eventually, but probably not HLL4 as it was originally planned.

conclusions after testing some random stuff:
- Great results with Lion - it learns very fast and generalizes well if you use bs >=64, low encoder LR, high decay as recommended by authors, but it likes to fry everything at low batch sizes. 8-bit Lion should be even faster. With low batch sizes using old adamw is safer
- Looks like most efficient way to train a model like HLL is to chain 2-3 trainings: 1: high bs low res with lion to learn general features and style, 2: low bs higher res with adam to improve details.
- Model learns from other AI-generated images faster than from "real" images?
- Adding just 10% of thread archives as reg images is enough to teach model how to imitate AOM/pesto look. Should be easy to get more interesting '"default style" with different reg images.

Navigation
View posts[+24][+48][+96]