If you like what I do, you can support me here-- https://boosty.to/snzhkhd/donate
Version 1 fix was tested and I didn't notice any artifacts, it's clean enough, so I decided to share it with you. I will try to make it even better in the future.
I would also like to thank https://civarchive.com/user/GustavGalla
This model can generate hentai without Lora, however, Lora can enhance and improve. Some anime lora works well, some doesn't (you can try to shrink all lora and add tags on realism and quality).
Description
FAQ
Comments (6)
How to use the checkpoint (V1.0 fixed)
I did a bit of testing. Look into the gallery, i uploaded a few testing images for reference.
Feel free to add points or correct me if i made a mistake.
Here are the results:
Sampler - Pick the one you like
Most samplers work fine - DPM Samplers or DDMI seem to get the best quality ones. I myself use DPM++2M Karras or DDMI.
(DDMI is a bit faster in return loses a bit quality)
Steps - 20 to 40 Steps is the sweetspot
<20: Images vary strong in quality and content. Not recommended in this range.
20 - 40: Images stay the same in content and get improved quality with higher steps.
>40: Images stay mostly the same with very little quality improvements
hires.fix Samplers - R-ESRGAN 4x+ works great
I tested which hirex.fix samplers work good (but i just tested like 3 or 4).
R-ESRGAN 4x+ and R-ESRGAN 4x+ Anime6B resulted in very good images. I recommend them.
(The Anime6B Version is more for anime based images but had not much difference to the normal version)
Hires.fix Denoising Strength - 0.0 to 0.5
When upscaling with hires.fix keep the denoising strenght between 0.0 and 0.5. Everything above 0.5 end in distortion and blurry images with no reference to the written prompts.
Prompting - Use enhancing prompts
Ensure to put enhancing prompts into the positive and negative prompts. I use
Positive Prompts: best quality, highest quality, masterpiece, high resolution, clean picture, smooth picture, high details,
Negative Prompts: worst quality, low quality, bad artist, distorted proportions
Content - Humans work, Furry maybe
Humans or similar beings (Elf, Succubus, ...) work very well. Furrys or similar need a Lora to work or maybe very good prompting.
Hands - Use a negative embedding
Hands may be the only downside so far in this checkpoint. They are not made from hell, just a bit distorted so its possible to work with.
Negative embeddings can help out like...
Bad-Hands-5 (https://civitai.com/models/116230/bad-hands-5)
Other embeddings work too (the ones below) but not as good as Bad-Hands-5
Negative Embeddings - Use simple ones
Negative embeddings can help to improve the overall quality of the images (if needed) - i think they are not needed, because the images are good by themself.
However i tested a few and the ones which do not impact so much on the image (a slight improvement if you like) worked better than the impactful ones which added noise to the image.
I recommend to use one of these (if needed/wanted)...
Boring_e621 (https://civitai.com/models/87781/boringe621-negative-embedding-enhance-images-stylistically-and-topically)
EasyNegative (https://civitai.com/models/7808/easynegative)
Image Sharpener (https://civitai.com/models/43286/image-sharpener)
The embeddings which added to much noise were...
Deep Negative V1.x (https://civitai.com/models/4629/deep-negative-v1x)
veryBadImageNegative (https://civitai.com/models/11772/verybadimagenegative)
Note:
Dont prompt full bodys when using low resolution, they get distorted - but this occures on the majority of checkpoints.
V1.0 fixed gives more freedom than the older versions. For example it was very hard to erase nudity like when you dont want a cleavage but a "closed" shirt. Now this works very easy.
Can you upload the older versions to HF?
@hagane_no_kintama They're still here, I just hid them. I'll open them now.
@snzhkhd thank you very much :D
is there a preferred VAE for this?
I use the basic vae-ft-mse-840000-ema-pruned .I have not tested others