If you like more cartoon-y anime art styles and huge breasts, then this LoRA is for you.
Since Hatomame is mainly an NSFW artist/circle, the model does best with NSFW images, but is also able to generate SFW images. There is some character bias, but you can get rid of it with well chosen negative prompts when it occurs.
If you like the LoRA, please leave a reaction, review and/or tip. Enjoy!
Usage
No activation/trigger needed. Simply include the LoRA in the prompt.
Recommended to use at strength 0.8 - 1.0 (start at 0.9)
Use Clip skip=1 or Clip skip=2 (other values may also work, use at own risk).
Works best at 1024x1024. Struggles to get the style right at lower resolutions.
Recommend to use with SDXL VAE or flatpiecexlVAE (depending on image content and preferences). Getting the color balance right may be a little tricky using only a VAE, but can easily be done using a decent image editing software (such as GIMP).
Example
See preview images.
Unpruned vs. Pruned vs. Lite
I offer three variations of the model, which only differ in the file size and quality. Pick the one that best suits your preferences. If you can not decide, simply pick the "pruned" version. Note that the quality is relative, even the lite version usually produces images of good quality.
unpruned: The originally trained model. Largest file size, best quality. Pick if you want the best possible quality, at the cost of disk space.
pruned: Downsized model. Almost same quality as the unpruned version, with noticably smaller file size.
lite: Drastically downsized model. Smallest file size. Still has good quality, but noticeably worse than unpruned and pruned.
Notes
This LoRA is quite the jump in art style, when compared to the Belko LoRA I did previously. Though both of them do well for paizuri content. Which is probably the reason I made them.
Starting with this LoRA, I also decided to release three variations for each version of my LoRAs (unpruned, pruned, lite), since it's pretty easy to make them and I think more LoRA creators should put the time (like 5 to 10min) into resizing their LoRAs in order to save people some disk space. That shit ain't free.
Next LoRA is gonna be Zheng, so probably less paizuri and more girls packing a little a lot of extra heat, if you catch my drift.
Using training data
Feel free to contact me if you want access to the training data.
The training data (images and tags) has been sourced from rule34.xxx and cleaned up manually by me. When using the training data or modified versions of it to create models or datasets, please credit both me and the original artist.
Disclaimer
This LoRA is not affiliated with the artist/circle it is based on. If you like the art style, please support the original artist/circle.
I could not find an appropriate socials link for Hatomame (aka. unlimited star), so I suggest you simply google them, if you want to see more of their work. They also make decent H-games, which you can find on DLsite.
If you are aware of an appropriate socials link, please DM it to me!
Description
This is the lite version. It has a minor, but noticeable loss in quality compared to the unpruned version.
Use this version if you want to save some disk space.
FAQ
Comments (5)
I was looking for this art style in Pony. Thank you!
It's not great. Only works with score_9, which reduces quality, and focuses too much on Ago's older, crappier style.
Addendum: Using "medium hair" in the prompt pushes the lora towards the newer art style. "Tomboy" does the same. It's something, but not ideal...
Hey, thank you for the feedback! You are right about the bias towards medium hair and tomboys (specifically: towards a certain huge breasted, medium brown haired tomboy). This is because it is hard to find varied images of their newer style. I added some older images to add variation to the dataset and remove some of the bias, but this has caused the shift towards the older style you mention. It's a trade-off. If you have a source for varied images of their newer style, I'd be happy to revisit this model and create a v2 which gets the newer style more consistent.
I could not replicate the issue with score_9. For me, using score_9 is usually beneficial to the image quality and not using it reduces the image quality. But the LoRA works either way for me. Could you provide some more details and maybe an example for this?
@StevesAISlop I'm not sure if civitai loses its shit if you post links so I'll tell you names instead. Look up [Hatomame (Ago)] 残陀村復興記 , [Hatomame (Ago)] BITCH FAMILY ON THE VILLAGE, and [Hatomame (Ago)] BITCH FAMILY ON THE BEACH. Those are all CG sets from their games so modern style, no text to clean up.
You can also look for Uchi no Futari 3, S.S.S -Shasei Support Staff- and Paizuri wa Sekai Kibo no Sport ni Narimashita, which are all in color but do have text. Luckily the text is often to the sides of the scene so it might be easy to deal with. Uchi no Shimai no Shakkin Hensai Force of Gigant also has a more modern style but it is black and white so maybe not as good.
If you have sad panda access you can find all of those there.
Lastly, he does have a couple more pics on both twitter and pixiv, goes by ago810 in both, but they are mostly content from the CG sets I already mentioned.
As for, score_9, it IS good but it's better to use the rest too because otherwise you lose a lot of detail, especially on the background. I've found score_9, score_8_up, score_7_up, seems to be the sweet spot where everything still works as it should. In my experience, this lora only REALLY kicks in when you use score 9 by itself, which is not ideal. I just uploaded a handful of images here on civitai for you to see, can get to them from my profile.
Having said that, I'm not even sure what would be the best practices for tagging when making a lora. Right from the get-go most lora seem to work when all the score tags are enabled so I imagine if anything it would be best to do that for compatibility and let the user decide how many they want to use.
@Detrian Thanks for the extensive suggestions! The CG datasets are a great source of images, although the variety is still a little lacking. But it should be enough to increase the reliability of the "newer" style.
I'm still not certain about what is causing the weird behavior with the score_9, since I never include any of these tags in the training data, because it risks the LoRA overwriting the original "meaning" of these tags too much. I might do some tests, but at this point I think it's just a weird interaction of the LoRA with the model, which may not even be present in the new model.
It might take me a little bit to swing around to this LoRA, since I currently have a lot of other LoRAs in the pipeline, but I can probably move it up in the queue a little.