Wake up anon, there's bugs to fuck.
LHC v0.4 is "done", by which I mean I am spending an hour per epoch with basically unnoticable improvements at this point, so since it is already better than the other versions imo, I'm calling it here.
https://huggingface.co/Jyrrata/LHC_XL/blob/main/alpha/v04/LHC_alphav04-vpred.safetensorsLora extract as well:
https://huggingface.co/Jyrrata/LHC_XL/blob/main/alpha/v04/lhc_04_extract.safetensors Generally very "at home" and probably works best with noob v-pred 0.6, haven't tested others yet. Much more heavy character bleeding than the model though not for all characters. Some work quite well
Lessons for the next version:
While previously I overtrained the TE, I think I could have used more/longer TE training in this version, cause some characters that the previous versions learned without issue are pretty "at home" in this version. Overall, undertraining is preferable to overtraining though I think.
Finetuning individual characters in the end stages can sometimes work, but is so unreliable and quickly undoes knowledge of other characters that it isn't really worth it.
I have grossly underestimated how long this needs to be trained for, and doing additional runs over and over again works, but is probably much slower than doing it in a single run. Even now, I'd say this is somewhat undertrained, but it has reached a stage where doing more epochs at a high LR would undo most of the later epochs that learned the finer details, meaning I'd basically be starting over again.
General tips for prompting: characters first, adding 1girl and/or solo helps keep out additional people, for related tags look at danbooru for those that have images there.
I will try to do an image for every chuuba at some point so that there's an example of working tags, this might take a while though. For now, I'll keep tagging more chuubas and retagging some of the problem chuubas. And also finally enjoy the model and do more prompting again. Haven't had the time to test SWG and the IP adapter yet.
Also: Am I just not remembering, or have we had a lot more schizo melties in the last few weeks/months than before? Is it just cause it's more active again?
>>90965416Thanks. Yeah, i hold off on scraping everything in media tabs and such, cause I'd end up throwing away 99% anyway. Causes me to miss some stuff though
>>90979074You know, the "uh oh stinky" monkey?
>>91119364Cute UI
>SquchanNoob actually already knows her, but to prevent her from being forgotten, I added her, yeah. It can do alheitrakzoth now though, which noob couldn't, at least not on its own:
https://files.catbox.moe/bz4x9t.png>>91132838Pretty cool. Couldn't find the lora, so I gotta ask, is that a realism tuned lora, or does it just work with the realistic checkpoint. If the latter, that's impressive at how well the lora adapts to the style.
>>91154170Yeah, darkness seems better, gonna have to test this to see if it's worth it to switch from 0.6. 0.65s was worse imo, so we'll have to see if the merging was the issue or not.
>>91154774Cute gg
>>91155376>futafags are always lusting after cockAfraid of a little pegging, are we?
>>91155873Afternoon TWL
