ModelsLab/[SD1.5] OpenSolera - Anime Mix - OpenSolera-α4 [Noto]

IMPORTANT ANNOUNCEMENT: In October 3rd, 2024, this model was found to have no VAE. If you downloaded this model before or on this date, or have the model with AutoV2 hash of 02ad770661, re-download the model. The fixed version should have an AutoV2 checksum of 29ee0a603a

Internal name: OpenSolera-a4-Noto.safetensors

This version is a re-merge based on existing models (with more added models to better align with my preferred style). I also moved my merge and testing over to reForge + ComfyUI (as LoRA merging on reForge doesn't work and SD.Next is too slow for me).

Also, Noto became this model's codename because I'm both bored and tired when making this. Maybe I'll start giving codenames (like what Cetus-Mix did) to make it less boring. The reason I chose Noto at the time is simple: Google's universal font that can display almost everything - Noto

Changelog + Recommendations:

• Added new models:

  • Checkpoints:

    • Nyan Mix - 230303_absurd (CivitAI)

    • FoxyaNSFW - v4 (CivitAI)

  • LoRAs:

    • (Changed) Ken Sugimori style LoRA - v3 offset - ChatGPT caption (CivitAI)

    • (New) Better Half-closed Eyes LyCORIS (CivitAI)

• Remade merge recipe with Triple Sum and less self-merging. Should be more reproducible while keeping most of the style of the first 3.

• More detailed and semi-realistic output. Alpha 3 was a good model on its own, but it wasn't the style I was looking for.

• No real need for positive quality tags anymore. The model is tested without it (on reForge)

• CLIP Skip 1 and 2 are both fully usable, each one with its own little differences.

  • CLIP Skip 1: More details, more saturated colors, (maybe?) better prompt adherence

  • CLIP Skip 2: More vibrant colors, less details, (maybe?) more pleasing images

• Mainly tested on 3:2 aspect ratios (mostly vertical)

• IMPORTANT: Using Anything's tagging style (tag everything you want to see) is still heavily recommended. Also, Hi.res fix (or equivalent) is a necessity for not-blurry images.

Specific recommendations:

• Furry/e621 tags users: e621 tags (anthro, feral,...) are semi-recognized and working (surprisingly) well. Furry models can finally be represented properly in this model, hooray! However, it only works in tandem with the proper species prompts and (sometimes) throwing "human" in negatives. Also, feral doesn't work as well as one might expect.

• LoRA users: No comment, this model should work fine with almost anything.

• Hololive gens: Heavily limited to what Hololive-6.3 has (No HoloJustice and anything after HoloX/3rd gen ID), especially in clothing. For other members, use LoRAs for them.

• Monster girls: Some monsters need their tags to be stronger for good results. Also, alraunes still don't work properly (and neither does AIOMonsterGirl extracted LoRA when generating alraunes). Also, since e621 works, you may need to add "human" to your images to keep as much human skin as possible, or go all in on making furry monsters with them (note that some are incompatible with furry tags).

• Other anime characters: Most popular characters should be recognized (as of late 2023 - early 2024). Blue Archive characters has a halo problem (halos are incorrectly generated, if at all), but the character themselves should be recognizable without much effort.

• NSFW content: 60% of the time it works every time... This is trial and error territory for complex stuff, but simple stuff should work fine.

• Artist tags: Untested. Use at your own discretion.