If you enjoyed this model, why don't you check out my other works? I put a lot of effort into correctly recreating artist styles and making the occasional character LoRA.😊
This model was trained using the collab/guide by holostrawberry [Guide] Make your own Loras, easy and free - colabs | Stable Diffusion Other | Civitai
About the artist:
Youtube: Speedoru
Twitter: 🤠 (@speedoru) / Twitter
Newgrounds: Speedo
Speedoru makes cartoony art and is most known for making animated music videos. They do not stop there but also create a lot of incredible and very horny art.
Weights & Prompts
The activation prompt is: speedoru_artist
For inspiration check the examples. I heavily recommend experimenting with the steps as well. Sometimes a 12 step image is just what you want, sometimes it takes 30 to get the necessary detail.
v2:
This should be your goto model. Use a weight of 1.0, adjust to taste. Check the XYZ plot in the examples.
v1:
Use a weight of 0.9-1.0, to taste. While the prompt is just "speedoru_artist", including "toon \(style\)" massively changes the image generation, getting much closer to the intended style. As usual I have included XYZ plots for reference.
Backgrounds and landscapes are cartoony with vibrant colors but less details.
Please leave a review if you enjoyed this model or have problems with it. I'd also be interested in how the model behaves with other base checkpoints, so why don't you give it a try and throw some images down below! 👇
Description
Trained on NAI with a learning rate of 2e-4
FAQ
Comments (13)
You got discord, I can think of a few art styles I'd like to see you try.
You can just throw them in the comments. Post a danbooru/gelbooru link or the name/website of an artist and I'll take it from there (no guarantees).
If you don't wanna post it here it's probably not something for Civit...
@StickyRicky I was thinking the Disgaea artstyle. Lots of character LoRAs for it, but no style LoRAs.
@CuppaGi From a quick google it might be a bit difficult to setup a proper dataset. Not sure where I'm gonna get 50-100 images with little text or game elements from. I'd guess you mean the cutscene art?
@StickyRicky Cutscenes, character profiles, stuff like that. I could probably put together a dataset, but how would I send it your way?
@CuppaGi Imgur or google drive. If they are public, thats extra info for people who are interested in making style LoRA ^^
Do you have the dataset you used available online on hugging or gdrive?
I'm sorry to say that I don't feel comfortable with openly sharing the dataset. The model creation is already in a grey zone and I used content from behind paywalls.
Basically it's just a mix of images you can find on newgrounds, gelbooru, patreon with a dash of screenshots from some videos (nsfw and sfw).
In any case I'll be reworking this model. It has come to my attention that AOM and NAI are inherently incompatible, so this model is probably a whole lot worse for all the people not using with with AOM.
@StickyRicky Would you be willing to do a Tutorial on how you train styles?
I would love to hear your thoughts / opinions on dataset collection and captioning - you produced a LORA significantly better than mine - would love to learn.
@DreamExplorer Honestly the guide at the top says everything you need to know.
I'm personally having a lot of success with this specific training rate, model type (LoCon) and model size (always going through about 400-500 images per gen. 100-125 images with 4 repetitions seems to work well). There is probably still room for improvement, especially with autotagged images and the variables I mentioned (still gotta try to manually add a "signature" tag to filter them out). Once I've done more testing on this (gonna take a while cause I have quite the backlog) I might create a small tutorial.
Some parts I didn't notice or wish I had known sooner:
Train on NAI unless you want to make the model just for yourself!
Set text_encoder_lr to 0 for styles.
You need a wide range of characters with different backgrounds and poses. One problem I had with a recent model was, that the artist only drew male characters with dark skintones -> it was close to impossible to translate that to female characters.
Finally the testing is the most important part. There is (almost) always a generation that works. Try them with simple prompts, prompts from the dataset and other LoRA to get a feel for their effectiveness and quirks. (Gen 25 for example was completely different from 24,26)
You basically want to find the strongest version that still allows you to effectively customize the output. That tends to be somewhere around 30 gens with my approach.
Hope that helps! 😎
@StickyRicky I really appreciate your guidance here.. any rules of thumb for captioning that you follow?
I've tried several guides.. sometimes I try to be very accurate in captioning and sometimes I under caption.. and sometimes just a single trigger word.
I've heard all different strategies when it comes to training 'styles' vs a character.
Any tips or tricks you have.. or any guides you found useful?
Thank you for reading.
@DreamExplorer So far all the tagging was done automatically using waifu diffusion
SmilingWolf/wd-v1-4-swinv2-tagger-v2 · Hugging Face
That's also part of the dataset maker on collab.
I just started manually adding prompts to exclude some things that it can struggle to detect ("multiple views" is a perfect example which I assume really helped with my Abbb model 2.0)
My experience with automatig tags ist that less is more but anything it finds will often get multiple fitting tags. For example, if there is a red kimono in the image, the txt will contain: "japanese clothing", "kimono" and "red kimono".
Don't add things that are technically there and visible but aren't visible enough to be detected as a difference from other training images -> connecting the prompt to it.
(The last part is a reasonable conjecture - haven't tested putting random prompts in there)
@StickyRicky Amazing response and so quickly. Thank you !
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.