Chloe Grace Moretz is probably most known for her role in Kick-Ass but since then she has been in several movies and shows including Carrie, Let Me In, If I Stay, Peripheral
Description
Since the model is made using custom merge, the simple prompts may return dull results. It's best to be creative and play with the strength of the token keywords.
Merge included HassanBlend and Pyros which means that there should be less anatomically incorrect poses (and yes, you could use it in the way HassanBlend and Pyros were trained for as well).
FAQ
Comments (16)
Scruffynerf is adding the same comment to all checkpoints trained on people, he's wrong. An embedding is NO WHERE NEAR as good as a fully trained CKPT. An embedding adds no new data to a CKPT, only helps SD generate better images from what it already has. A full CKPT trained with new images produces far better results. It's up to you, don't DL it if you don't want to waste space but you will get better images from these.
Do you know how to train a ckpt with images and its respective description?
i'm using base shivamshirao which can't do that
you would need another tool (for example automatic1111, which interestingly also uses shivam, but can use descriptions)
I was thinking of making both a trained model and an embedding with the same images, that way the model can be merged if someone wants, then use the embedding to help counter any possible diluting from the merge.
I extracted a lora from this. Do you mind if I release it for others, giving you credit of course.
@ryoko2 if you said it a week ago i would say yes, but today is the day i'm realising the loras of my models ;-)
but thanks for taking interest in it :-)
i will post a link to it soon!
and thanks for asking first :-)
@ryoko2 here is the Lora version: https://civitai.com/models/8244/chloe-grace-moretz
please check and if you feel like your extraction produces better results - post it too (and let me know how you did it :P)
also, i am curious what base model did you use since most of my celebrity models were not trained on plain 1.5 but a custom (never release) mix of several models.
@malcolmrey I would say the results are comparable. I used koala_ss and used its extract lora feature and the 1.5 as the base model as I did not know what you used. Left network dimension at the default 128
I'm going to do a review in the lora and add comparison pics of the 2
@ryoko2 i use kohya_ss (you probably wrote the same but got autocorrect? :P), also 128 dimensions
i trained my models on a merge that i did in november/december and never released that merge (it was only a couple of models, all based on 1.5)
it's actually interesting that it works for you nicely, i guess the lora is just probably bigger because it contains not only Chloe but the difference between the merge and 1.5
I just stumbled on this convo... I realise people do steal shit all the time but this is a great showcase of how open source communities are meant to work.
@samwyse yup, and i love it!
open source for this kind of stuff is the way to go :)
better remake it like a lora model
I'm not sure what you mean, I already converted it to LORA model :)
https://civitai.com/models/8244/chloe-grace-moretz
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.