Obligatory Disclaimer: This LoRA is intended to be used alongside style LoRAs and/or artist style tags to bring out its full potential. All example images use WAI-Anima and the "haku89" artist style tag.
If you don't see all 5 example media, make sure you're logged in and try changing Browsing Level.
Comments and gallery posts are encouraged and appreciated regardless of their content! (Within ToS, of course.)
Start with a model strength of 1.0. If you find the model to be too stiff, try lowering it to 0.80.
A very quick and dirty character LoRA for the character Donna from the indie visual novel game Best Friends For Dinner! Trained on a dataset of only 20 images that consists largely of cut up screenshots, images generated/enhanced by ChatGPT, and one art.
Full transparency: This model really isn't meant to recreate the original style of the original artwork from the game. At the moment, this isn't really realistically feasible due to the fact that pretty much ALL of the available images of this character are in the exact same pose just with different facial expressions. If I were to train using only these images, there could be some issues regarding overtrained angles and poses. Hell, even with the dataset I used, the model still has a predisposition toward the downward looking angle seen in the original VN artwork.
Due to the lack of available training data, this model is far from perfect. It's a quick and dirty model made in less than an hour because I really liked the character design. Use or download with tempered expectations. The lack of an official full-body reference image forced me to assume what anything below her dress looked like. Due to the simplicity of the design, I assumed that she was simply barefoot and had Gemini whip up a quick full-body reference. I went with human feet over sheep hooves both because hooves aren't my area of expertise and personal preference. Likely due to the limited dataset, the artstyle can be weirdly inconsistent at times.
Using tags like cartoony artstyle, cartoon, flat colors, simple shading, and thick black outlines can help bring out results closest to the original artstyle should you want to recreate it. That being said, I recommend using specialized checkpoints and LoRAs to more strongly imitate the original.
Ultimately, not very glaring issues in practice but still something to keep in mind. The below and attached triggers are general suggestions. You can almost certainly get better and more consistent results by more specific prompting. I'm just lazy. Play around with it to see what works for you. Remember, Anima loves natural language so converting the listed triggers into natural language can result in more consistent generation.
Detailers are highly recommended for the face and eyes. Especially the eyes since she has abnormal pupils.
Manual inpainting or other image refinement methods are recommended if you're looking to accurately portray small details.
I will most likely update this LoRA if any more official images or fanart of the character that can be used as training data are posted. Especially in regards to full-body images.
All example images were generated without a refiner step. You can probably make your generations look much better with improved workflows.
Triggers:
BFFDDonna, snow white skin, long hair, ahoge, white hair, blue eyes, large breasts, sheep ears, sheep tail,
black collar with small bell, blue overalls dress,