BlipMix - variety of styles at high quality, very good base for training LoRAs

The model is a result of various merge techniques (DARE merge, extract merge, and simple add merges) on primarily Pony based models, alongside some more versatile SDXL models. BlipMix is one of many merges I play around with locally. It's high enough quality I figured it's worth sharing.

The model can produce a variety of styles at high quality, using pony based tags. Example prompts are in the showcase images, with plenty more open for discovery.

LoRAs:

I've found it to be excellent as a base model for training LoRAs. Perhaps due to the diversity of style knowledge in the model, it picks up on styles/artists/concepts very well. I'll be posting some soon.

Involved models:

autismMixConfetti
sdxlNuclearGeneralPurpose

wildCardXLv4

uncannyAF.ponyXL

ZonkeyRealism4.2
Plus possibly more I forgot to document...

Method:

The exact method originally used not even I know. I'm just a mad alchemist mixing potions. I can confidently say though that confetti was the largest part of the merge for v1, with the others DARE and extract merged with first confetti then subsequent merges at smaller scales. The result is a model capable of many styles including photorealism as well as ponyDiffusion's more "fun" aspects.

I will post all my stuff for free, including other LoRAs/checkpoints and coding tools like useful scripts/webapps, such as a blind rating elo ranker for LoRAs. You can find them here.

If you'd like to support me, every little bit helps! As I said I will post everything freely, but living do be costing money. Feel free to buy me a coffee, even a dollar helps ;)

描述:

训练词语: