Verified: a year ago
SafeTensor
Type | |
Stats | 696 |
Reviews | (62) |
Published | Jun 17, 2023 |
Base Model | |
Training | Steps: 1,950 Epochs: 30 |
Trigger Words | Etie, orange hair croptop, long sleeves, floral print, bowtie, frills, skirt, ribbon belt, thighhighs |
Hash | AutoV2 CE23EC65B3 |
Etie trained at 1024x1024. CAUTION: Generating at 512x512, 512,768 etc may not always produce the best results. Mainly just wanted to do this for my favorites.
V2: use at weight 0.8-1
V1: use at weight 0.6-0.8
I think the quality increase is pretty nice, but doesn't always play well with "pose, etc" LoRAs that require certain dimensions. I'm mainly using the Tiled Diffusion/VAE to gen at 768x1024 then upscale with hires 1.5x. Took the 16th epoch.
Will still release, 512 trained as usual, but might do this given I got time (took an hour to train instead of ~10 min). Also trying out new training method with cosine+restarts, instead of constant. Also read Jiwenji's tensorboard article, which was useful.