Read "About this version" to see what changes were made to the model. I might make changes you don't like and you may want to stay on the older version.
The only authorized generation service outside Civitai is yodayo.com
Maintaining a stable diffusion model is very resource-burning. Please consider supporting me via Ko-fi.
AingDiffusion will ALWAYS BE FREE.
EXP models will be updated here to reduce confusion: https://civarchive.com/models/52780.
===
AingDiffusion (read: Ah-eeng Diffusion) is a merge of a bunch of anime models. This model is capable of generating high-quality anime images.
The word "aing" came from informal Sundanese; it means "I" or "My". The name represents that this model produces images relevant to my taste.
Guide to generating good images with this model
(NOT REQUIRED SINCE v7.7. FOR AINGDIFFUSION v7.7 AND UP, SET THE VAE TO NONE)
Use the VAE I included with the model. To set up VAE,you can refer to this guide.Use any negative textual inversion (e.g. EasyNegative), they will help you a lot.
Recommended samplers are "Euler a and DPM++ 2M Karras" for AingDiffusion v7.1 and up.
Hi-res fix is a must if you want to generate high-quality and high-resolution images. For the upscaler, I highly recommend 4x-UltraMix Balanced or 4x-AnimeSharp.
Set Clip skip to 2 [optional, if you need more creativity to the output and not following the prompt 100%], ENSD (eta noise seed delta) to 31337 [to replicate image], and eta (noise multiplier) for ancestral samplers to 0.667.
FAQ
Q: What's up with the frequent updates?
A: AingDiffusion is a model I use daily, not something I merge to gain popularity or for the sake of download count. I made constant efforts to improve the model whenever possible and wanted to share the improvements as quickly as possible.
Q: I can't generate good images with your model.
A: The first thing to remember is that every little change matters in the world of Stable Diffusion. Try adjusting your prompt, using different sampling methods, adding or reducing sampling steps, or adjusting the CFG scale.
Keep experimenting and have fun with the models! 😄
Description
I HOPE I DIDN'T MAKE ANOTHER BLUNDER THIS TIME
Reverting things back to Anime screencap-look by default.
Trained a whole new dataset for this update, with corrected epoch, steps and everything. Should be better than v10.5 by a lot.
===
I currently do not have any volunteers to test my models, unlike a few months ago when I had volunteers, so now I often judge the model performance by myself and by showing a couple of friends the output images. If you're interested to be my volunteer beta-tester and giving feedback, please email me at [email protected], or DM me on X (formerly Twitter) @kayfahaarukku.
FAQ
Comments (3)
Version 11 seems like it makes every prompt to look kinda teensy and somewhat it gives a "chibi" likeness to anything. Very different from V10 and 9
The new custom dataset is kinda biased tbh. You can (kinda) fix this by putting "mature female" to the prompt. No guarantee, but it helps a little bit.
The other thing is "1girl, 2girls, etc" is really a biased tag on booru (having chibi, mature female and child mixed into this tag). This also a factor that make the dataset captioning not very good (since I don't manually tag all the images, all automated). Try to avoid these tags and use something like "solo, mature female".
nice