An Azuriade Zenith

AnAZ is a mix of my old model mix and 5 more recent models : 3 anime, 1 more digital illustration oriented to add more special effects and details, and 1 more comics-oriented for better separation between character and background, better outlines if asked for, and a slightly different colorful color palette.
(my old model was a mix of 4 anime models, including 1 NSFW-oriented one)

Results are mostly anime, and become more 2.5d with corresponding tags.

Since this model mixes both anime and non-anime models, it works with both CLIP sentences/descriptions, and Danbooru-style tags. The best would probably be a mix of both, so feel free to try new combinations!

My goal here is to make a model that would help me make more detailed anime characters and backgrounds for my NSFW stories, while keeping the faces in the style I like.

And if that could help you get closer to the style you were looking for, then that would be great too!

If you like it, please consider buying me a... tea! https://ko-fi.com/anneazuriade

*'¨``¨'*¤•-An-•¤*'¨``¨'*¤•-az-•¤*'¨``¨'*

Known possibilities :

• Anime portraits of humans, elves, catgirls/nekomimi and other kemonomimi or monster girls

• Medieval fantasy and modern styles; also futuristic to a lesser extent

• Landscapes, with or without characters

• Action scenes (though it will require many test batches, tweaking options, and cleaning/inpainting)

• Furries too, but won't always work well or not on the full character(s). I'll make another model mix for that. Half-furries might work though.

• Still recognizes a wide range of famous anime/game character names. Their style is more diluted than in other models, so you might need specific LoRAs to improve the results. This model is better at making new original characters.

*'¨``¨'*¤•-an-•¤*'¨``¨'*¤•-Az-•¤*'¨``¨'*

Technical details :

This is my 3rd attempt at merging models, so I'm not sure of the extent of its possibilities, as well as I might not be aware of all the issues. Please send me images and/or comparisons if you see any weird!

Known issues :

• This v1 might have inherited too much contrast and 2.5d effects compared to what I was looking for, though I'm still in the process of re-learning how to make good prompts with this mix.

• Without quality tags or with low resolution, the eyes can become weird; with color tags, they may get heterochromia (different colors). These didn't happen as often in models used for this merge. If it happens, try increasing CFG and/or steps, remove VAE if it's still on, or increase denoising in highres fix.

• Slight increase in results having 3 arms, unless you use corresponding negative tags or embeddings against that. Changing a few weights on tags or changing CFG/steps will remove the 3rd arm most of the time.

• Like in many old models, long hair or fluffy kemonomimi tails might merge with background details. If it happens, try changing a few weights, increase steps or CFG, or add more tags detailing the parts that blend together.

• the first 6-7 example pictures were made with a VAE activated (anythingVAE) which deteriorated the quality and some colors. See the VAE section below for details.

*'¨``¨'*¤•-an-•¤*'¨``¨'*¤•-aZ-•¤*'¨``¨'*

Recommendations :
Most example images were made using :

Sampler & steps: DPM++ 2M Karras starts giving good results at 20 steps but 25+ is recommended. DPM++ SDE Karras starts being good at 25+ steps. Euler and DPM++ 2S a Karras both give correct but eerie results at 15 and more normal ones at 30+. DPM++ SDE needs 30 steps minimum and results can still be eerie.
DO NOT USE with this model : Euler a, DPM++ 2M SDE, DPM++ 2M SDE Karras, UniPC. These give completely broken results (either noise instead of face, or whole picture becomes noise or foggy blur)
Also avoid ancestral samplers (with an 'a') unless you need them specifically. They are usually slower due to how they work.
In complex pictures with lot of details, the result can sometimes be improved by increasing steps up to 90, after which there are more meaningless detail changes than gains.

Resolution: works well up to 1280px width for landscapes and up to 1024px height for portraits.
-> If you get extra limbs or elongated ones, try reducing resolution in the direction the issue is happening.
-> If you get duplicate characters but want to keep a high resolution, try adding corresponding tags : 'alone', 'solo', '1boy' or '1girl'. In negative : 'two', 'brothers' or 'sisters', 'group', '2boys' or '2girls', 'men' or 'women', 'couple', etc... Adding more background description will also help removing unwanted additional characters.

CFG Scale: 7 by default. Up to 13 may increase contrast, details and fix the eyes a bit. Once you found a good face/pose/scenery, try changing the CFG to see if it can improve the result.
If you need to go above 15, try changing your prompt instead, switching from description to Danbooru-style tags (without the _ underscores) or vice-versa (check example images' prompts for a few ideas). If any of these fail, you may need to use an image as a base, or use ControlNet, Regional Prompting.

Highres fix: R-ESRGAN 4x with 20 steps for landscapes and background and more realistic-looking 2.5d characters, R-ESRGAN 4x+Anime6b with 10 steps for anime (2d) characters and/or without complex background. Denoising 0,25 if the image already looks perfect, otherwise 0,4 will hopefully fix minor details, but might make the eyes smaller and the character look a year or two older.

Clip Skip: 1 for maximum detail; it that's too much detail on characters, or if the model's style appears too strong in a particular prompt, then try CLIP skip 2.

VAE: Don't add a VAE; the model was trained on models that already included a VAE. It's probably kl-f8-anime2 since using it doesn't change much to the picture. Anything or NAI VAEs will tone down the contrast and colors, so not recommended.

Positive prompt: add "masterpiece, best quality, intricate detail" at the end of your prompt; check examples for additional ideas.
Negatives: use a "bad hand" embeddings, then you can try adding: "worst quality, low quality, monochrome, greyscale, amateur, (blurry), (text), error, jpeg artifacts, signature, watermark, username, ugly, (bad hands)"