Future note: This model is a legacy model and is only for reference. For better performance, check OpenSolera.

Solera - A "jack-of-all-trades, master of one" Stable Diffusion 1.5 anime model

From Wikipedia: "Solera is a process for aging liquids such as wine, beer, vinegar, and brandy, by fractional blending in such a way that the finished product is a mixture of ages, with the average age gradually increasing as the process continues over many years."

I made a bunch of merges with the same name before, labelled it different versions to differentiate while I learn to merge models on my own, but none of them were uploaded. This is one I'm satisfied with enough to release it to the public. It will probably get buried by better foundations like SDXL, but I'd rather have it out there so I can finally complete this journey.

What is this model?

It's a merged Stable Diffusion 1.5 model, with the following models used for the merges:

• Kakigori V3 (https://huggingface.co/nubby/Kakigori)

• Cetus-Mix - CodaEdition (/model/6755?modelVersionId=20229)

• Pastel-Mix - Full (/model/5414/pastel-mix-stylized-anime-model?modelVersionId=6223)

• YiffyMix - v37 (/model/3671?modelVersionId=512554)

• Counterfeit - v3.0 (/model/4468/counterfeit-v30)

• Koji - v2.1 (/model/41916/koji)

What is it good for? How do you use this model?

It's a general use, 2D, anime-styled model that is meant to be used with minimal amounts of negative prompting while still delivering decent quality most of the time. It can do both SFW and NSFW images, leaning more towards NSFW.

This model mainly uses booru-styled tags (refer to the Danbooru Tag Groups wiki for more details on tags and Koji's "Using the model" section for more details on usage), but can also use natural language when needed.

Note: Do not use underscored tags (like "red_eyes") when generating with this model, use spaces instead.

I mainly tested this with DDIM, DPM Adaptive, Euler a and DPM++ 2M Karras (or AYS). Basically, most of them will work well enough, but I like the looks of DDIM most.

This model is also mainly tested with SD.Next's "Full attention" parser (see here for more info) or ComfyUI's default parser. Other parser haven't been tested as thoroughly yet, so please post in the comments if you use them.

Standard quality prompts (e.g: best quality, absurdres, masterpiece,...) works well. Negative quality prompts (e.g: worst quality, low quality,...) works decently as well. View the images for a better understanding of this model.

Recommended resolutions: 512x768, 768x512.

You probably should pair this model with inpainting for best quality. (FYI: none of the preview images have been inpainted or adjusted in any way)

Upscaler choice: 2x AniScale ESRGAN or 4x Lollypop.

CLIP Skip is tested at 1, while 2 is untested but functional.

The model already have a baked-in VAE (from Kakigori), so no need to grab another VAE from elsewhere.

What is it bad at?

Like most other Stable Diffusion 1.5 models, hands, legs (in different positions), and text are a nightmare for this model. I don't know much about text, but DWPose works decently well with this model to fix it.

Also, as it is an anime-styled model, realistic images are not a goal for this model. Realism is not the goal for this model.

Prompt adherence with this model is more on the "hit-or-miss" side, leaning towards "miss" depending on what you're trying to do. The model (when generating in 768x512) sometimes misses an entire leg.

What happened to SoleraV1-V6?

The first set of models (SoleraV1 to SoleraV3) are completely lost to time. They were simple merges anyways, not worth uploading here. V4 is on my system, but it's not yet good enough in my opinion. V5 took on too much of PastelMix, so I barely use it. V6 (internally named SoleraRemix) was an attempt to create a new base for Solera to work on, but I still felt something was missing. V7 is the current iteration of the mix and the one I feel good enough to upload to CivitAI for reviews.

What else should I know?

If you like this model, go and try the original models I linked above. Without them, this model would've been impossible to make.

This model may not be used on any paid generation services. You may not sell this model or any merges that contain it. This model is for personal/private use. Commercial use is not authorized.