SD1.5-OpenSolera

This repository is meant to be a backup for my Stable Diffusion 1.5 models in case they ever went down on every other site I have published them on.

What is this model?

For people familiar with my previous model: This model is a remake of my original Solera model for commercial use, based on many of the models used in it and additional models for more functionalities. Currently, a6 [Fleur] is the last version of the model until conditions are better. Until then, a1 and a4 are decent starting points, with a4 being closest in style

For people unfamiliar with my previous model: This is a merged Stable Diffusion 1.5 model meant for generating 2D-2.5D style anime pictures made from the following models:

Checkpoints:

AuroraONE 
Something v2.2 
VividOrangeMix - NSFW 
Hassaku - v1.3 
YiffyMix - v37 
Hardcore Hentai - 1.3 
Mistoon_Anime - v3.0 
(a2 onward) ElixirProject - v1.6
(a2 onward) Nitro-Diffusion 
(a3 onward) Furryrock - v_7.0
(a4 onward) Nyan Mix - 230303_absurd
(a4 onward) FoxyaNSFW - v4 
(a6 onward) Himitsu
(a6 onward) HorridHentaiMix 
(a6 onward) True Futa - v1.5
(a6 onward) PornMaster Anime - V6 
(a6 onward) IdentityCrisis - EroDemonMix_03ws 
(a6 onward) AbyssFluffyKemono

LoRAs:

AIOMonsterGirls - LoRA/LyCORIS extract
CluelessC's Hololive - v6.3-eps-resized-512 
(a3) Ken Sugimori style LoRA - v3_offset 
(a4 onward) Ken Sugimori style LoRA - v3 offset - ChatGPT caption 
(a4 onward) Better Half-closed Eyes LyCORIS

The main goal of this model is to create a "jack-of-all-trades, master of one" anime model that can do almost anything. This model can output SFW and NSFW images, but current testing indicates that this model tends towards SFW more, with decent NSFW outputs.

All models have a baked-in VAE, Check the model details for what VAE was used.

How to use this model?

Prompting

Same as most other anime models, this model mainly uses booru-style prompts with no underscores (see Danbooru's Tag Groups wiki for more details and NovelAI's Basics page alongside looking at the tags and AI metadata on booru sites (e.g: AIBooru, Gelbooru, Danbooru,...) for usage), but can also use natural language in some cases.

When prompting this model, be specific and mention every detail you want to include. Basically, prompt it like Anything V5.

Also note the prompt parser used as well. This model is mainly tested against SD.Next's "Full attention" parser (as of a3), reForge's defaults (from a4 onward) and ComfyUI's default parser.

This model can generate good results using minimal positive quality prompts and negative quality prompts. To get a better idea on this, look at the example images.

Positive (optional): best quality, absurdres

Negative: (worst quality:1.4), (low quality:1.4) + any additional issues

Note: Most of the strengths listed here are suggestions. Do what you like, I don't take responsibility for your bad prompting or what you generate with this thing.

Recommendation: Do not use any negative embeddings with this model. Not only does it completely destroy the style that it has, it also makes prompting accuracy plummets off a cliff.

Sampler

This model is mainly tested against DPM++ 2M Karras, but DDIM, Euler ancestral, UniPC (with ComfyUI's "simple" scheduler) are other great options.

Step count is dependent on sampler, but 20 steps is a good start for most. For furry generations, lower steps works better.

CLIP Skip works best at 1-2, depending on each version.

Resolution

This model supports all three normal SD1.5 resolutions: 512x512, 512x768, 768x512

Each resolution deals with things differently, with prompt accuracy being best at 512x512, followed by 512x768, then 768x512. Details-wise, 768x512 is the best (imo) and 512x768 can generate at similar quality.

Post-processing

For best images, Hi-res fix (or latent upscaling) is recommended. Current recommendation is bicubic 2x upscaling.

For model upscaling, these are my personal recommendations:

x2: AniScale2 Omni + Refiner, x2 AniSD SPAN 190k + DeBleed

x4: 4x Remacri and x4-AnimeSharp

Q&A - Lightning Round:

Which model should I use? This or the original Solera?

Use this one. The original Solera has a couple of differences, but not significant. For legacy and testing purposes, I still keep SoleraV7 around for comparison. In most cases, I think this model would out-perform my original by quite a bit, if you can tolerate losing some contrast.

Why OpenSolera?

It's Solera, but open to everyone (including commercial uses). Pretty simple, really. Also differentiate it from the original Solera model. Internally, this model has another name: SoleraV9.

What happened to previous versions of Solera?

V1-V3: Lost to time. They're all very rudimentary merges of AbyssOrangeMix3 (specifically AOM3, Anything V5, and Hassaku.

V4: Don't like it that much, still exists but not really usable or useful.

V5: Absorbed too much of Pastel-Mix, completely destroyed the background. Was my main driver in 2023.

V6 (internal name SoleraRemix): Don't like it that much. It was a decent foundation for future merge ideas, but not much remains.

V7: Uploaded to CivitAI as SoleraV7-a9.

V8: Went nowhere. Like, actually nowhere. It was the same situation as V6

V9: Uploaded to CivitAi as OpenSolera. Incremental updates are marked as "Alpha" versions

V10: Not officially called SoleraV10 by me, but Alpha 7 [Sake] is basically that. The OpenSolera line continues from a4 onwards with new codename-based versioning. No more versions of OpenSolera-alpha will be released until I can think of traditional ways to improve this model.

Where did the name "Solera" came from?

From the Infinity Bottle: "An infinity bottle, also referred to as a "solera", "fractional", or "living" bottle, is your own personal blend of whiskeys that is ever changing. By simply combining a bit of your favorite whiskeys in an empty bottle (or by adding to a base whiskey) you get a personal, never-ending blend of your own creation. This blend will likely have never been possessed or enjoyed by anyone else, ever. As you continue to add more, the flavor profile of the bottle is always new, evolving, and taking on a life of its own."

I don't actually drink myself, but I find the concept interesting, so I adapted the name for my merges.

Where's the merge recipe?

Check the zip file labelled "Training Data" on the version you're downloading. The actual merge takes too long to list now.

Basically, it's 13+ steps with a lot of long fractional numbers. Can't really copy them over properly.

Version 6 and onwards are merged using ComfyUI, so their merge recipe is included within the model itself.

Do you actually use this model or are you just making them for clout?

Yes, I do use the models I create for making things (see my Chub.ai page for an example, all images comes from a version of Solera (see above) or OpenSolera for examples). I rarely generate things, what with lack of motivation or ideas. I don't want to post all the time, only when I want to.

Is this similar to SD-Silicon?

No. This model series does not use any form of automation (except XYZ Grid for testing) in merging. Goal-wise, though, they are quite similar.

Is this similar to @AnonymousM Basedxx series of models?

Weirdly enough, yes. It's not an attempt to clone their work or anything, but I did get inspired by them when making this series of models (especially in using the Hololive LoRA). Other than that, I'm not sure if there's any other similarities.