>>12517 128:64, I haven't tried 1:1 but I've settled on this as it seems to work.

>>12519 fair enough, I think I want to try and figure out a lower dim to bake at, might go a bit higher than needed then resize down with dynamic resizing. I feel like the bake issue is probably more of a dataset issue rather than dim size issue though, because most of the details are there. If this 4th bake doesn't turn out well, then I'm gonna probably try and prune the dataset, remove any images without the halo. hopefully it doesn't completely destroy the balance I had between shun and shunny, not that I can't just change repeats around to make it work

>>12520 You can always resize the lora with the other script after you do the bake so its' better to start high imo. I haven't tried these after resizing them either but that'll be the next step once I'm fully comfortable with the ones I got since fuck having this file size on these. I've got 150gb of lora's and most are just various rebakes of shit because I can't decide what's best.

>>12516 >I've been having the biggest aneurysm with Iori's because nobody draws her in anything but her swimsuit and her halo is like 5 different fucking colors in fanart, been losing my shit over it. Time to get into the turbo autism that is manual data correction.

>>12521 my experience so far is that you can usually fit everything into dim16 usually, if I bake at 32, then I should have plenty of headroom to learn everything then dynamically resize down. to be honest though... I haven't baked at dim128 in so long that I literally forgot how

>>12523 I should test lower stuff but it takes me like 1.5 hours per lora to bake, and my PC is fairly unuseable except having a stream open so it's boring if I don't do it either sleeping or at work. I've settled on 128 for now since it doesn't seem bad just the size is off putting, but really want ot make new stuff I haven't really made new stuff in a while just keep re-baking and re-tagging older shit to get it better. It's working so it doesn't feel like time wasted but it's feeling like a grind.

>>12524 oh that's 100% fair, with torch 2.1.0 though, my bake times have cut by a lot. I just baked 2400 steps, it took 1 hour and 20 minutes, usually that shit is like 2 and a half hours. So at this point, I'm willing to try and bake at lower dims to get it correct, that being said, it's not entirely correct, but the halo is actually decently consistent now. seems like training for longer at a lower lr did the trick in this case. it's still not entirely correct though, so I'm gonna prune the dataset and try again later probably, might also just change to dim32 to get it a bit easier

How many multi-concept/costume folder LORA makers do we have here? I'm wondering how you balance stuff for the best results

>>12525 I'm still hesitant to set up torch 2.1.0 so I'm kinda waiting for it to get folded into automatic1111, not sure if it will but I should look into sooner then later since more people then not seem to get far better speeds.

>>12527 it's not hard to do desu, all you have to do is change the torch line in the launch.py file with >pip install torch2.1.0.dev20230320+cu118 torchvision0.16.0.dev20230320+cu118 --extra-index-url https://download.pytorch.org/whl/nightly/cu118 and then run the webui after you delete the venv folder, it doesn't require its own xformers like sd-scripts so you can just use opt-attention

>>12527 as >>12528 said, it's not hard to install.

>>12526 Not an expert or anything, but basically you need the subsets relatively balanced so the model has ample time to learn from each instead of getting lazy with the smaller ones.

>>12528 remember to use code tags [code][/code] or else it'll look messed up like that

>>12531 oh I didn't know that, thanks

>>12496 You can uoh in peace here brother

>>12530 On a related note, trying to gen 3 girls at once is proving to be a shitshow even with ControlNet and/or masked Latent Couple and/or multisubject LoRA like penis-on-face. I'm not sure if repeating the data of the 3 of them together, while not enough to generate consistent gens on its own, would help those techniques. Because 2girl sex gens do much better even though there isn't substantially more data and mostly non-sex. 1024x1024 is very expensive for iterative experimentation... I also want to test LoCon...

So is there any reccomended method or concept to finetune? I know HLL4 is diverse because there's so many vtubers and different artists so the model is an ideal finetune however I just want to know when do I want/need to finetune compared to making a LORA

Went back and trained an embed for the first time in a while. How in the fuck is voldy/monkey patch's embed training so much slower than Kohya's? It's like several orders of magnitude slower, Kohya's embed trainer can finish in about 30 minutes something that takes monkey patch a whole day. I can't believe it used to be in common use.

>>12535 Hll was always trained with at least 100k images split between chubbas and porn/generic. 3.1 was 75k chubbas + 150k generic/nsfw. He hasn’t posted his full spec yet on Hll4 beta but did drop a txt file with his tags. He does not do that much tag management other than checking for blatantly wrong tags and mostly let’s the auto tagger do the work. When I started my Fate Zero/Ufotable finetune I started with 38k images but results started looking better at 90k images and and hitting the 110k started making stable looking generations. I’m currently training a model with 116k ufotable + 30k Troyca/El Melloi Case Files + 34k assorted porn images with some more in-depth tagging autism because WD-tagger will do some of the weirdest fucking shit when tagging TV resolution. Hope that gives you an idea if you ever touch the finetune game. I had to start looking at other places for info because other than hll anon, no one bothers with training anything but LoRAs.

>>12536 I was told Monkeypatch + 4090 gave you 30 minute bakes with embeds when the Monkeypatch update came out in late January.

>>12537 Maybe I can make a finetune on 100k idolmaster pictures, can work... man Hydrus is going to be busy soon

>>12539 If you want it to do porn you are gonna need additional generic/nsfw images but for a test run, 100k should be enough to get you started. What type of images are you planning to work with, and what GPu do you have? Me and Hll anón have 4090s and 100k images with the right settings will take about 18 hours to train.

Dunno if anyone's using it, but I updated https://rentry.org/anonskohyaentrypoint to also support training embeds, and made it so you can queue up various different operations at once in the same queue.

>>12464 >>12465 >>12467 Looks like the same problems still remain from previous version, still have to specify hairstyle and eye color etc i guess ill download it and try with the lora on top

>>12540 two 3090s, no NVLINK yet cause my stupid motherboard didn't come with SLI support.

>>12543 But before that will the learning rates be vastly different? Given that finetuning requires more VRAM I assume that batch sizes are going to be like 4-5 right?

OHAYOU! I find it deeply fascinating that this locon shits the bed as soon as you gen something bigger than 512x512

>>12546 how is that the case? Did the train the LORA under 512x or something?

>>12543 You the anon that just upgraded his PC? I believe you can still run multi GPU training without NVlink.

>>12547 weirdest example

>>12548 Yeah I just wanted NVlink to have something cool, but true honestly I don't even use up all my vram and can still play games with my training settings it's only really when I do over 1000x res for LOHA/LOCON and batch 3 that the VRAM starts to be used up fully kek

>>12549 or here first one is 512x512 second is 768x768

>>12551 unironically pickled, where did you get this? 5chan?

>>12498 literally me. I think I only generated like 20 images of Nahida after I perfected the eyes

>>12552 https://huggingface.co/KoRo8888/Ohayou_Locon

>>12554 literally wondering what the hell they did to even get a LORA to get weird shit like that