The updated LoHa is in bundle format and it can only be used on webui version >= 1.7.0
See https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13568
Alternatively you can download from
https://huggingface.co/alea31415/Frieren_Beyond_Journeys_End/blob/main/EP01-10_LoHad8_clip2/frieren_EP01to10-loha-30000.zip which contains lora and embeddings separately.
Please use both embeddings and LoHa in this case to ensure the effectiveness of the model. For more details on pivotal tuning please see https://civarchive.com/articles/2494/making-better-loras-with-pivotal-tuning
All the checkpoints can be found at https://huggingface.co/alea31415/Frieren_Beyond_Journeys_End
You may want to use different checkpoints depending on your goal and model. For example, Eisen is hard to learn so using higher epoch-checkpoint can be more effective.
Wildcard:
Note for AuraF, FernF, StarkF, and EisenF (You can change embedding names yourself if you first unbundle than bundle)
AuraF
AuraF, elbow gloves
Draht
Draht, white shirt, suspenders
EisenF
EisenF, beard, horned helmet, armors, red cape
FernF
FernF, long sleeves
FernF little girl
FernF little girl, long sleeves, hair bow, red bow
Flamme
Flamme, choker
Frieren
Frieren, elf, earrings
Granat
Granat, white shirt, stubble, epaulettes
Heiter
Heiter, glasses
Heiter_old
Heiter_old, glasses
Himmel
Himmel, cape
Himmel_old
Himmel_old, closed eyes
Linie
Linie, hair ribbon, dress, black ribbon, neck ribbon
Lugner
Mazoku_xx
StarkF
Macht
Macht, half-closed eyes, hair slicked back, long sleeves
screenshots
fanartThis model uses my updated pipeline https://civarchive.com/articles/2383 which constructs the dataset automatically from anime.
Description
The first model is kind of overfitted so I retrained the model for better flexibility.
All the screenshots are trained with 'screenshots' (4852)
All the fan arts are trained with 'fanart' (125)
However this also means that we may need to compromise the character fidelity sometimes (especially for eisen, heiter, himmold, heitold etc). Including "screenshots" in prompt enhances fidelity but also brings the anime style.
Changes
1. dim 32 -> 8
2. tte lr: 2e-4 -> 4e-5
3. clip skip 2 -> 1
4. base: nai -> ACertainty
5. improve no character screenshot dataset
The training of Eisen is delicate. Even after pruning 'beard' and 'mustache' it still gets basically associated with 'horned helmet' and prompting 'beard' may still be needed for better resemblance.
P.S. Himmold and Heitold are for their old version. I get rid of such design of trigger word with pivotal tuning later.