For all the fans of satisfying cake cutting videos and "Is It 🎂?"
The LORA model was trained very quickly on quite low VRAM (24GB) locally. Even though the dataset was in a low quality and it was only roughly a 3000 steps training, this turned out much better than I anticipated. Even to the point where I deem retraining unnecessary.
I use HNoise and LNoise both on 1.0 strength.
Be descriptive!
This is the most important part - it is crucial to say exactly what you want to happen and ideally to give this LORA some details to work with - what happens when you cut it open, how the object behaves - is it spongy? is it not? - etc.
This model is intended to only be used in the object cutting + cake reveal context. For any other usage I can't guarantee any good results😉
Example of a very simple, but effective prompt:
c4k3LX, a knife cuts the shoe, it's surface layers peeling off and act like a soft cake under knife's pressure, revealing it's interior to be a soft multi-color layered cake.
For any questions, please ask away! I'm very new to this, it's first LORA I trained and I'm myself still learning a lot every time I try stuff like this out :)
Maybe more LORAs will be coming in the future ^^
Description
FAQ
Comments (12)
24gb is plenty for training Wan. Low VRAM is 3060 style.
3060, 12gb, 32gb system RAM. Dual mode, single file for low and high.
That's low VRAM, but still doable and worth it for lots of basic LoRAs.
The "training data" zip file has everything you would need to know to replicate the technique using musubi-tuner.
https://civitai.com/models/2165860/goth-girls-style-single-file-375mb?modelVersionId=2438999
I'm sorry, i have never trained a lora and feel pretty lost with the topic, I have never gotten a totally clear answer about this from anyone so I am asking you who seems to know his stuff:
I have a 4080 Super and 64GB system RAM, can I train wan loras with it? If so, do you know a guide or something on how to train it with only 16 GB VRAM? I think I would have some great ideas for concepts that would not just be the same thing a million times like it currently is but realistically I don't have the time to learn it properly without a straightforward guide. Thank you
Edit: I realize and already have realized while I was writing this comment that the information I'm looking for is actually in there, but I thought who knows, maybe you have any other guide you know of that is really good and comprehensive. Sorry for being too quick
@Radyschen I'd say it's possible and what I've followed was one by the Ostris for AI-Toolkit, which is his own creation and what I used to train this LORA. You can take a look at his youtube, he has a guide for Wan2.2, but some new settings he added since that video have made it even more manageable :)
@Radyschen - you totally can, but with lower specs you start to hit some real limitations. Some concepts could be complex enough to require voluminous data... which would extend training times on low vram cards and older cards to unreasonable durations.
But for almost anything you might want to train, your 16gb 4080 is actually better than any of my current GPUs ... and I train the shit out of wan loras ...
I can hold your hand in PMs if you need it, but I will also elaborate anywhere.
@Elix_LX do u have the link
please don't zip the models, then lora manager cannot track it and get metadata etc. upload high/low in separate tabs please 🙏
I decided to go for it since I somehow got errors when uploading both HN and LN models when creating the post. I added them separately now, lmk if it works :)
@Elix_LX thank you! yes it works fine now
@Elix_LX Civitai just does that sometimes when uploading. You have to try over a lot, and/or just try again later.
@WithoutOrdinary Sometimes? Always slow and sticks
@qek Yeah, I lied. CIvitai's upload almost always sucks for many of us. I gave up uploading stuff a year ago, I waited 2 years for them to fix the issues and there was no progress on that or any of the dozens of other accepted issues I had reported. xD
how long does it take