Type | |
Stats | 1,165 |
Reviews | (65) |
Published | Nov 24, 2023 |
Base Model | |
Training | Steps: 4,480 Epochs: 4 |
Usage Tips | Clip Skip: 2 Strength: 0.8 |
Trigger Words | dkttlr |
Hash | AutoV2 6C119CE534 |
If you like my contributions to the community, please consider Buying Me A Coffee or check out my other sites on AllMyLinks
RECOMMENDED SETTINGS
CFG SCALE IMPORTANT - Use CFG of 3-5. Anything over 5 might suck.
PROMPTS - "Smiling", "from behind" and others are trained into the LoRA. For a full list of the prompts/captioned used for training, check the Lora file's metadata in your webgui
TXT2IMG - A weight of 0.7-0.9 will give the best mix of likeness and flexibility.
INPAINTING - A weight of 0.75-85 will give the best low-noise likeness. If you want the absolute best likeness, switch your checkpoint to SD15 since that's the model on which it was trained.
FACE RESTORATION - isn't usually necessary, but I recommend setting your face restoration strength inside auto1111 settings to something like 0.7-0.8 which makes the effect less than the default setting of 0.5.
TRAINING INFO
v1 & v2 were trained on 128 photos (face & full body)
v3 & v4 were trained on 224 photos (face & full body) and reproduces both with more detail and accuracyv1, v2 and v3 were trained at 768 resolution, so 768x768 inpainting works great
v4 was trained at 1024 resolution, so 1024x1024 inpainting works great
GETTING THE BEST LIKENESS WITH LORA'S
The checkpoint you use can heavily influence the likeness of a celebrity Lora. I train my Lora's on SD15, so you'll get the best likeness on any model that used SD15 as the base model. See this example for how a checkpoint can influence the look of a Lora.