santa hat
deerdeer nosedeer glow
Sign In

_Envy_'s Cauldron 01: Mixing Tips


(Edit: I didn't actually mean for the cover image to be just a picture of boobs. It was cropped that way. Anyway, it's funny, so I'm going to leave it.)

There's a lot of great, very specific information out there about mixing and particularly training, so what I'm going to do instead is try to give more general tips for producing stable models and refining your own training and mixing methods.

Anyway, keeping it brief for now, here are a few tips you can use to get stable mixes that work well with lots of different models (note that I'm saying LoRA here, but this all also applies to lycos):

  • When mixing LoRAs, get a feel for which ones are stable and which ones aren't. Mixing stables LoRAs will generally result in outputs that are stable as well, whereas if you throw an unstable one in there, it can ruin your whole mix. I'm going to call one of my own LoRAs out here as an example of something unstable. Envy Elegance has what I would consider to be a really nice effect, and I would mix it into most of my loras if it didn't cause things to get unstable really quickly. Because of that, it stays out of the mixing cauldron (although I will often use it in addition to other LoRAs to spice up my images).

  • If you're mixing loras that you haven't already verified to be very stable, mix them in at low strengths. A lot of my mixes are a combination of a few of my own that I know to be very stable at strength 1, plus a number of others at strength 0.2 to 0.5.

  • Mixing a lot of loras at low strength can give you a stable output. You can mix 8+ things at once with great results if they're stable enough.

  • If you're using the kohya_ss GUI, watch the command line when you mix things together (it outputs the command line in the console), and you can learn how to run it directly from the command line, which gives you more options. The GUI really wants to normalize your total mix strength to 1, and in practice I've found that this is almost never what I want to do, and would result in a very weaksauce mix. The command line allows you to specify individual strengths for each model you're mixing, and mix more than 4 at once.

  • When you find a good set of loras to mix together, try randomizing their weights a bit and see if you come up with anything better.

Feel free to ask questions about mixing here, and I'll try to tell you what I know. My next Mixing Cauldron post will be about training LoRAs.