Welcome to C4ttitude!
C4ttitude is part of the "Attitude" series which will feature more animals and concepts in the future. This is only the beginning ^^.
C4ttitude is a LORA custom trained on cats images and artworks that will make easier (and way more fun) the cats + wearables creation in Stable Diffusion. This version features a flat color/line comic art style AND a renowed flexible but less stylized ver. We also have another version in our Patreon featuring a more painterly style (adding comparison at the end).
Which is better? To be honest I liked both and wanted to release this concept for free ^^.
Version 1:
It works with "c4ttitude" at first part of the prompt and then you can select color of the cat, class, etc. Mind that sometimes the color of the cat will influence the color of the outfit if you don't have enable cutoff but you can do very good outputs just with normal Stable Diffusion settings.
Version 2:
Refer to the changes in the version.
Prompt like this : <lora:name:weight> portrait/fullbody portrait/..., coloryouwant c4ttitude, outfit, places, etc. Rest is normal. c4ttitude is intended to replace cat bias cause "tails". This version is way way more flexible but maybe less stylized in some seeds. Choose what you want!
IMPORTANT!
I wrote an article here and it's like my finetunning diary for the ones interested with releases upcoming, trainings done, exclusive patreon models previews, personal thoughts, etc. So if you want to see what I'm planning or what I'm into, feel free to take a look.
Enjoy and support us so we can keep creating quality content:
https://www.patreon.com/user?u=90661954
Thanks for the support and stay creative!
Description
Flatter ver
FAQ
Comments (14)
Very nice. Is there a way to get color without stripes?
Umm, sadly not. Just tested right now different checkpoints, clip skip 1, striped lines in negative and even lowering to 0.2-0.3. It's impossible cause all datasets had striped cats. Didn't think on this when doing dataset, drawings, etc. Sorry.
That's ok, it's still nice :)
@n_u_l_l Thank you, glad you like it ^^. Open to ideas, etc.
Curious about something, so the whole training set consists of striped cats, was "striped" or "tabby" included in the description of every image? (I'd do that to make sure it's trained as a separate concept and separatable)
Great lora and great choice! These cats are cute :)
@z0d You're right, all set were striped except 2-3 photos of black cats. I trained v1 as "c4ttitude, cat, color, etc." and was not the best choice cause AI understand it as double meaning (c4ttitude vs cat). Maybe not in all generations but yes in some. One of the v2 partial fixes I did (overpainting tails in Photoshop aside) was tag "portrait/fullbody. c4ttitude color, etc." so this way I have room for fixing. Tabby cat was never in the tag, same as striped. Being said that, don't forget that even its not on the tag AI's behaviour is to learn from text/tag plus how it sees the whole thing while batching/repeats.
I'm experiencing human-related prompt terms really pulling the generation away from the concept the lora produces, like 1girl etc.
Sorry I think I don't understand, you mean you are prompting "human terms" with this LORA right? You don't have to put human/1boy/1girl, etc (in fact you shouldn' here). You need to put cat and c4ttitude. Other way the cat factor dissappears
I like it. Unfortunately im getting multiple Tails in almost every image. 2-3 in one image xD Colorful V3 model.
Did you use easynegative? Maybe using too much negs. Can cause that and same for >1 value. Will try at night...I found some but not so many. Dreamshaper 3.32 clipfix
@LDWorksDavid hmmm... yes, maybe too much negatives. I will try in next days.
@AddictiveFuture Don't worry mate. Testing.
Edit: You were right. Gonna make some fixing on the dataset and retrain a v2 to see if I can fix this.
Edit 2: Released flexible ver. if you can try will be great. Not promise amnything but seems to work better. Read the info changes ;).
@LDWorksDavid V2 seems to be much better :)
@AddictiveFuture thanks. Commented just that, avoiding cat word and replacing c4ttitude to it (ex: tabby c4tittude, black c4ttitude, etc... now understand c4ttitude as an specific new cat bias and knows it has 1 tail, haha) in v2 fix almost every problem of native bias. In v1 fixed some problems but not all. In the future will try to see if I can maintain more the style of v1 and reprocessing it. 😄