Liara T'Soni from Mass Effect series by Bioware.
Works better with lower CFG vaules (4-6).
Recommended weight range is around 0.4-0.7 depending on the model and CFG vaule. I found that between 0.54-0.6 works decent enough with HiRes fix.
Weight could be reduced when using SD upscale or certain img2img actions.
Example promts; masseffectliara standing in a room / close up of masseffectliara standing in a room / portrait of masseffectliara.
Extra Info:
Dataset for the model was around 30 pictures including pictures from the web, in game and manually captioning images.
Released as an alpha, i will try to improve it further on my free time.
Crediting when using this model in merges and such is much appreciated.
If you want to support my work, buy me a coffee!
Description
Initial Release.
FAQ
Comments (9)
FINALLY!!!
nice finally a decent Asari/Liara. you know what they say once you go blue nothing else will do.
They just never tried turians, not some generic recolored human females.
if all goes well maybe a samara , Aria or Benezia?
Seem a little inflexible, it captures the likeness pretty well but can't seem to change outifts.
Yes, majority of the dataset includes Liara's default armor.
@Taloji The flexibility issue could be easily corrected by descibing in the caption files the armor/clothing worn in each individual image in the dataset (easily done in bulk using Notepad++ or equivalent).
Whatever you wish to be flexible in the finished lora, you WOULD describe in the caption files; backgrounds, clothing, lighting, surrounding objects (chairs, walls, cars, etc.), arm/leg placement, general action or position (sitting/standing/laying), camera angle, camera distance, style (painting, cartoon, cg render), basically anything in each image that is NOT your intended subject. These are things you typically want to be flexible and not appearing in every image generated by the finished lora to some degree. The lora then learns or recognizes these things but gives them a very low or virtually non-existent weight.
On the other hand, SD learns and applies much higher weight to anything NOT described in the captions when creating a lora or lycoris. For characters/people, you generally do NOT want to describe the face, body shape/weight/size, age, skin tone, eye color, hair color, etc. as these are all things that keep the character/person recognizable.
Of note on the last paragraph, adding very generalized facial expressions within each image to your captions will increase flexibility for emotions & expressions but you may lose a little character/person facial semblence depending on which checkpoint you're using with the finished lora. Just be certain to keep it general ("an excited look on her face") and do not specifically describe the eyes, mouth, etc. that make up the expression ("wide eyes, smiling mouth").
Without the proper use of captions during lora creation, the end-user has to work especially hard to negative prompt any undesired elements out of existence, and/or must lower the lora's weight to such a degree that SD can no longer accurately generate the intended character/person. It can also make using additional loras a nightmare. Hypothetical example, if I don't caption out the blue shirt my subject is wearing in each image and I then use an additional lora trained explicitly to apply red shirts, the resulting images generated by both loras together will be a train wreck no matter the weights or prompts I use.
Sorry for the book. I felt I would write this as it's always disappointing to see great looking loras that turn out to be so inflexible.
@Bit_Shifter Yes again, i know how to caption properly, i do have LoRA's with multiple outfits/expressions and such, this one was not one of them.
I'm uyikrtyj4ertgh and this is my favorite model on the civitdel