OpenSolera-α3

Internal name: OpenSolera-a3-bless09-rf7-1899-0041-3929.safetensors

Yes, this model has gone through Model Toolkit pruning already. "0041" is WD1.4-bless09 VAE, "3929" is the CLIP from the LoRA merge step.

Changelog:

• Additional models added: Furryrock v_7.0 (CivitAI) and Ken Sugimori style LoRA (CivitAI)

• Experimental CLIP replacing (see this article of mine for more info)

• Better multi-person prompts (faces are slightly less distorted at further distances, but Hi.res fix is still recommended)

• Maintained decent hand quality (at least the fuck-ups aren't as bad in my experience)

• Prompt adherence is... a bit worse? Negative prompts are more required than before, along with detailed prompts.

• Flatter base style than Alpha 2. For those looking to mimic Alpha 2's base style, use "flat color" in negative prompts (see example image).

• Better furry generation (mostly due to Furryrock). It's still not great, but it's getting better... I guess.

As usual, for people wanting to check out the model's merge recipe, download the "Training Data" zip file.

Important side note: I recommend using CLIP Skip 2 for this model, as it's the closest to what I intend the base style to be like (with "flat color" in negatives).