i don't know if this site can handle tucker :)

you have outdated lycoris extension, you need to update it from the extensions tab: click on the "check for updates" and then update the extension "a1111-sd-webui-locon"

@Santaonholidays yes, two ways

1. go to extensions/a1111-sd-webui-locon and run this command from console: git pull

OR

delete extensions/a1111-sd-webui-locon, run the webui and install this extension again

@Santaonholidays yeah, that is the one

do you have the A1111 itself also updated to rather recent version?

i am on version from February 20 (65995a2e f71a3c9c) and it is good enough

@Santaonholidays another person experiences a problem (albeit a different one?) and I've suggested switching to the same version and uninstalling additional plugins that could interfere, here is the discussion:

https://civitai.com/models/35792?commentId=96016&modal=commentThread

If that does not help, might be a good idea to make a thread on the webui github about it (or take a peek there as someone might have already reported it)

--

what I can tell is that when I first found out LyCORIS I looked around and found the extension, downloaded it, and that worked fine;

but when I started doing my own LyCORISes I had some errors - I could not load them, and I had to update both the plugin and webui itself, fortunately, it worked but I'm just saying that you could probably have a similar problem where there is a temporary version mismatch

Thank you :)

I have left some comments on several model pages (started in Agnieszka Dygant model page), but you are right - might be a good idea to make a rentry page :)

@LeMarquisDeGucci yes, this is pretty much spot on, thanks!

the difference now is mainly that I extract it to LyCORIS and not to Lora (extraction tools are in this repo: https://github.com/KohakuBlueleaf/LyCORIS.git) and work great!

@LeMarquisDeGucci I would say two: flexibility and likeness

After extracting LORA from the DreamBooth model, that LORA worked rather well (but not perfectly) on the base model of that Dreambooth, but failed to give me decent results on other models (which was one of the main points of LORA!)

also, when I made several images with the same prompts using Lora and base model and then the dreambooth - I knew immediately which ones were done where meaning - the dreambooth had visibly better quality (which is understandable because youtube guides [such as nerdy rodent] were mentioning that LORA training is good but not as good as dreambooth)

I tried various parameters, I tried extracting only the difference (so only my training) and sometimes the training and the base model (so extracted against vanilla 1.5 and not the direct base model); I tried different network sizes (this helped a bit or maybe it was placebo, but the files ended up being 300MB) - it had little to no effect

when lycoris came about, I pretty much got better results (compared to lora) in my initial tests, I ended up using a smaller network for extraction so the size is below 100 MB yet the quality is much better than LORA.

as you can see my contact with Lora/LyCORIS/LoHA is not direct - I'm not training them as people usually do, I just do the extractions (and I haven't touched LoHA at all).

hope this helps somehow :)

I think the key is to not go with high CFG, I never go over 7.

What you can always do is get something that looks rather nice and then all the imperfections you can clear with inpainting.