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 for NAI using the collab/guide by holostrawberry [Guide] Make your own Loras, easy and free - colabs | Stable Diffusion Other | Civitai
The generation was selected based on testing in AOM3. I recommend experimenting with the weights when using other checkpoints.
Check out the artist:
Instagram: 美好よしみ (@yoshi_mi_yoshi) • Instagram photos and videos
Suzuri: 美好よしみ ( yoshi_mi_yoshi ) ∞ SUZURI
v1.1 and 1.05 are just one generation removed but I didn't even notice when mistakenly creating example images with the "wrong" model. There might be the occassional detail - in particular when combining it with character LoRA - that is better represented in one gen or the other.
Use a weight of 1.0, adjust to taste. As usual I have included an XYZ plot for reference.
Some useful prompts to get the right look:
"colour" theme, "colour" background, simple background, high contrastAs a sidenote: I messed up with the naming of the original project. The prompt is consistent with the LoCon name but still has an extra Mi.
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
110 images trained at 4 repetitions per generation with a learning rate of 2e-4.
FAQ
Comments (4)
Looks really good, thanks for taking my suggestion!
I don't have the time now, but I'll definitely mess around with this later today.
This painter's style I also train, but the results are not good, often overfitting, using the model ANYTHING.
Your training model is great, can you share some training ideas? I'd like to learn more, thank you very much!
I have never used the model ANYTHING. It looks a bit strong artstyle wise, which might conflict with making lightweight models.
In general try keeping the learning rate slow. This allows for a wider selection of acceptable generations. Then don't always expect the model to get the style 100% right. Sometimes specific prompts, like "flat color" or "toon \(style\)", will make all models a lot better.
To prevent overtraining you can always try to use it with flexible character LoRA. If, for example, the hair color changes you have gone to far. (Try the jinx LoRA or my riju model)
Finally, Lycoris tends to be a bit better than LoRA, but that is just an opinion.
Some other experiences from my side: You want a minimum of 50, if possible 100+ training images. Be careful to select recent ones, if you have a wide selection and don't just scrape them. Artstyles change after all. You could also limit yourself to the profiles and make a dedicated model - this would probably improve the head placement and line thickness a lot.
In the end, it's a whole lot of testing, testing, testing but keep at it. Experimentation is the key, because I sure as hell don't have the perfect method yet.
@StickyRicky So that's it, thank you for your answer, I've learned a lot, thank you very much!
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.