>>2133

>I don't think there's a way around it. The best option to avoid concept/object bleeding IMO is to keep your loras separate when adequate instead of having multiple concepts packed into one. You can see this with the debreasting lora (awesome btw, if creator is reading), where using debreast_fly/mutilate/carve separately actually does a cleaner job than using the multi-purpose one.

Yeah I'm suspecting the same. Since it already works so relatively well I wanted to try going the unified pussy stabbing -Lora route first and the compartmentalised aproach would have been my plan B. But now that I'm seeing those same thoughts I've had layed out by you here I'm beginning to think that separating the sword//knife types and training a Lora for each will most likely be the better way. I'll sort the dataset and see what's there and what's missing.

>Multidiffusers is an alternative upscale method similar to hires. It adds back and fixes a lot of detail and in my experience is more stable than other methods, although it does come with its issues when genning obscure stuff. But if what you got works fine for you, no need to change it.

Interesting, I'll give it a look.

>I always have Aya on the background at around 0.1-0.5, it's good support, specially with anytest, picrel

Also interesting, will try that out. Any triggerwords I need to add when using aya like that?

And let me guess, did you make that pic using one of the versions of the original for this? (Obviously I Added the hand and handle for the Lora-training)

>Sounds like you could use some optimizations. What UI are you on?

I'm using A1111. xformers enabled and --medvram. Also tried --medvram-sdxl but it felt a wee bit slower somehow. Im working with a 3060 Dual with 12gb vram, a maximus ranger VIII mainboard from 2017 with an i5-6600K and 32gb ram. (#So_ashamed xD)

So if you know any more startup cmds I should try I'd love to hear it! :)

>>2154 ouch! that hit the spot, I guess oO ...also thanks for the earworm! >.<