Krea2 v2.0:
So yeah I trained again, same data set, though this time i redid the captions with ai tool kit's auto caption.
side note training krea2 taxes the old 3090 a lot, depressing that 24gb of vram is not really enough these days.
Ok the example images, I used a few random prompts and some of the old prompts too, the workflow embedded is a lora testing work flow i created, so may not be of use to anyone. I do have a version of this lora that did better in some cases, but this one did the best over all.
For nsfw you will want to use a nsfw check point. My training data includes a few nudes, but its more about the facial structure and shape of the woman than anything "naughty".
Use responsibly, enjoy. As always its free... as all lora's should be!
May be the start of me making more....may be a one off, lets see how its received.
Z-image v1.0:
First lora in a LONG time, and first lora on Z image turbo,
I followed the video tutorial from the AItoolkit dev (Very cool tool!) and used my dataset from flux.
This makes women more cute, subjectively in flux it was made to combat flux chin.... and failed a bit at that, in this its more of a test, but I did notice womans faces all seemed to generate around late 20's early 30's so this dataset has more early 20's (and 18, 19) faces.
The trigger word is ukgirl (not sure if its needed or not)
Note: I have not had time to do more than a few generations, though hit no snags in my tests.
Description
updated to Krea2.
No triger word needed
FAQ
Comments (24)
How are you training locally on a 3090? I cant even start a task without a crash with a 4090.
Use ai-toolkit and select offloading. You can train with part of or the whole model on system memory. And the speed is not even that much slower.
@Loraholic I tried that, same error. Maybe because I only have 32gb RAM?
I'll order some more RAM, just need sell one of my cars...
@pogo Yes that is very possible. I have a specisl setup that require much more ram then usual when i train, so i have no idea how much you acturally need. I need 128GB myself to not crash with my setup.
OOM crashes at init usualy just dump all the vram and just stops,without any error. At least for me when i hit the limits.
use one trainer instead. Much faster than ai toolkit and much lower memory foot print. I'm training krea 2 on a 5070 TI right now.
@Goonald_Trump This comment makes no sense. The only way this is possible is if you use offloading to system ram as well, or you are training with quantization.
@Loraholic Couldn't get it to work :(
Maybe some day I'll get back to training LoRA for the community
Hi, I did not change any settings other than turning samples off, and offloading vram.
I used the Raw version of Krea2 to train as that is supposedly better for training even though we are all using turbo krea2 models.
To save ram, you could untick 1024 i forget where it is at mo, that will effect your training quality though.
This training was 2000 steps and took 2 hours... it taxed my 3090 vram to the max and took half or my ram (64gb) I see you mention having 32gb of ram....that is likely the issue. sorry chap, wish I had a magic bullet for you.
I do have a solution....of sorts.... runpod... no clue how to set up now, but i used it before and you can rent a nice gpu for $1 an hour or less if you look around the site. You will need to hunt tutorials on that.
I wish Ideogram had taken off... thats supposedly WAAAAY easier to train, but Krea2 is just so much more popular that ideogram has been mostly abandoned (shame)
Good luck!
@Loraholic Yeah of course, I didn't mean to imply I wasn't. I meant to @pogo addressing his concern of only have 32gb of system ram.
My current run is 25 images at 768 resolution and getting 3-4s/it. Using 16gb of vram and ~30gb of system ram. So he should be able to run with a 4090 (8 extra gigs of vram) and 32gbs of system ram.
@Goonald_Trump Is it as easy to set up and go with krea as ai trainer? I say this knowing fully well that AI Trainer, for Krea, has not been set up and go for me 😥
My training on a 3090 was very slow (50s/it) but it did work (I only tried the turbo model with training adapter so far), advisable to turn off all sample images (as the sampling froze up for me on RAW).
@J1B I heard training on the turbo model in ai tool kit (with adapter) resulted in bad lora's and told to use the krea2 raw version (which is what i used for this lora) Have i been misinformed, did the turbo model work well for training for you?
@AI_man2025 With respect, can you all stop talking about your experiences training in a thread I started by saying I can't train at all? 😭😭😭
@AI_man2025 Yeah, I had heard that as well, but it I couldn't get Training working on RAW locally ( I think I just have to disable Sample images though) so I haven't done a comparison but it still seemed to train pretty well on Turbo for what I wanted.
I just want to add some experience i have with turbo adapers and raw models. If you want to change the model completely with a totaly new style, use turbo adapter. But if you want to change the original model use raw.
I have always used raw for all my LoRAs, for both Z and Krea2.
But if you want to make something like "Melancholy" etc i think the turbo adapter is the way to go.
@pogo in reply to the post where you said maybe its your 32gb.
I am currently training another lora and I am using offload ram as it does go a bit faster on my rtx3090 (fill a gpu vram full and it slows down, or so i have found) Anyway...
21.8gb vram used
33.8 gb of ram (that fluctuates + and - 1gb)
I know thats not a good thing to hear, but it does show that in my very unprofesional, unscientific test, your ram is the issue.
Its in the tweaks, I am able to train a character lora on 10gb of vram over night, I use AI-toolkit
musubi uses dynamic fp8 and that divides vram cost while not sacrificing speed. 5060 ti has trained many local loras.
I have 32 RAM and a 16 VRAM 3080 Ti, but can train Loras with OneTrainer, let me know if you want a json file.
(the file https://pastebin.com/dqWU8zAj)
@foshizzlemanizzle I will probably give it a try if you want to send it. I'm just worried about investing time for low quality results. I'm so happy with the results I get from AI Toolkit, but I have to rent gpus instead of using my own.
I have a mobile 4090 (16gb vram) and 32gb ram, and I can comfortably run training (at least at 512 and 768 resolution) with around 30% transformer offload as well as cache text embeddings, training is quick and lora results are pretty great!
sidenote, increase your pagefile to 40+gb so that the training doesnt crash while its quantizing
@Doootboi are you talking about onetrainer or ai toolkit? I would love to make ai toolkit work. I've made so many great models with it in the past
@pogo yes I am talking about ai toolkit :) 32gb ram is very limiting so that pagefile increase is mandatory, though speed ends up being fine. You also have more vram so that should help a bit too