AAM XL Anime Mix
Anime Screencap Style Model
Do you like what I do? Consider supporting me on Patreon 🅿️ or feel free to buy me a coffee ☕.
A ❤️, a kind comment or a review is greatly appreciated.
Join my Discord Server
This is essentially derived from AAM AnyLoRA Anime Mix, but it's based on SDXL.
It won't work with loras and embeddings for SD1.5 even if they're trained on the original AAM or AnyLoRA. "Mix" as in mix of anime styles.
I suggest you use CFG 5-7 (not higher than 8), 20-30 steps with Euler a.
Upscalers suggestions are None (bicubic upscaling in comfyui), any good GAN for anime or Latent (only if you know what you're doing).
For Turbo I suggest you use CFG 3-4 and 8 steps Euler a or 15 steps LCM.
Unlike with DreamShaper Turbo, I think base AAM XL is better than AAM XL Turbo most of the time. However the latter is MUCH faster.
Purpose of this model
Make amazing anime style artworks on its own.
Train character loras.
Use anime styles.
Generate anime art and stylized art.
It can generate hentai on its own, but you might need loras and embeddings for specific stuff.
ComfyUI Workflow: https://pastebin.com/BHCEzc6T
Finetuned over "DreamShaper Anime", which is a mix of Anime Art Diffusion, Hassaku and DreamShaper XL. Finetuned using the AAM dataset. Fair AI Public License 1.0-SD
Description
Doesn't do unwanted nsfw and fights back nsfw prompts. Negligible quality loss compared to the base one.
For Turbo I suggest you use CFG 3-4 and 8 steps Euler a or 15 steps LCM.
FAQ
Comments (26)
Very expressive. The depiction of the human body is natural.
相当不错的模型，我呆了，太np了就是
I'm struggling with this model. After making many Loras with PonyXL where usually around 5-10 epochs is enough for most styles, I wanted to try a model that have less bias, well... 8 epoch is CLEARLY far from enough with AAM, the outputs are worse than without the Lora, the backgrounds are super-simplified, it feels like samples from mid-training of a SD1.5 model (see : neHqrSa.jpg (7168×1164) (imgur.com)).
Also I don't understand how come many people in this page are using 1.5 negative embeddings, this is SDXL. I'm lost.
With or without Loras, it's pretty difficult to get anything good, maybe I spent too much time on PonyXL ¯_ (ツ)_/¯
Does the model work with AUTOMATIC1111? Whenever I generate images, they end up looking multicolored (like a rainbow lol) and pixelated.
me2
I guess it's because of the Lora. The model does not support Lora that was created for SD 1.5. Could be possible that I still had negative promts that didn't work with this model when I tested it.
Lora or vae from SD
Me too, all my pictures turned into a bunch of pixels and dots
most likely your cfg value is too high
using on automatic1111 with no issues whatsoever
Does anyone here feel that the quality results created by this checkpoint decreased after Civitai updated its generator engine?
Yup. It definitely isn't doing the absolute knockout girls that I made a few months ago. Really sucks.
It feels like quality of generation decreased over time.
Great anime model, thanks.
Do you use a specific VAE with this model?
hi there i note one thing today,when i use img2img ,and i put trees,plant, put them in negative prompt ,they cause damage to the generated image, d u k why?
PEAK
Did anyone have solution for hand?, my result always messup at hand
For certain prompt or tag, this model will just produce randomly noise picture. It happened from time to time and just wondering if there's anyone else have the same experience?
text-to-image, WebUI
positive prompt: (best quality),(flat color,anime),hand
negative prompt: None
Sampler: Euler a
size: 1024*1024
Clip Skip: 2
seed: 1574082842 (doesn't matter, tried other seed too)
Outcome: https://imgur.com/O7YhEfs
What clip skip did you use?
@Zaranna clip skip=2 - tried clip skip=1 and still noise
Same problem here.
@Nley Just found a solution myself - remove all brackets
e.g. best quality,flat color,anime,hand
and the model works - not sure why though
Anyone tried taking this to 60 steps? and over CGF 7? (say 8.5?)
With 60 steps, background objects get much more fine detail, and patterns become more geometrically precise. There doesn't seem to be any sort of issue with overcooking. It's not particularly sensitive to CFG; it seems to mostly affect color saturation, and it starts to produce artifacts once it gets up into double digits.
I wish there was more SFW checkpoints. This is amazing. Please consider training more SFW artistic checkpoints
Nice