Wan2.1 text 2 video Lora
Actually, I train with the "press against the glass" concept, but somehow it also works with shots taken inside the glass wall or inside a car, as long as I use the right prompt.
Example prompt : Breast press against the glass in shower room , Ass press against the glass
I spent around $10 on Runpod to experiment with this LoRA, so I have to charge Buzz to fund building more LoRAs. Sorry about that.
Description
Text2Vid
FAQ
Comments (10)
I felt bummed that you felt you had to apologize for asking for a small amount of buzz temporarily. This stuff does cost money to train, in addition to time learning and setting up data, and often needs refining over time with later versions.
I think charging for early access is a perfect way to handle things. If someone is impatient, or just want to show you thanks, they can help you out, or get it for free later. I don't see how anyone can hold it against you.
$10 is actually cheaper than a lot of people's attempts. Could you share your training settings? I might start training some at that price. I was looking at $50 based on the training other loras have required.
Thanks for your heartwarming comment!
As for training, I'm honestly not very good at coding, but I found a WebUI interface here: https://www.runpod.io/console/explore/njiks0yigu
an easy-to-use interface that automatically downloads the base model. You just need to upload your dataset along with the caption.txt files.
After a bit of setup, you can start training right away. The downside is that it requires 48GB of VRAM to run this template.
I used the default settings for my training with a 256px dataset (I’m not sure why, but someone mentioned that using smaller image sizes is fine).
@pathAi Ah yes. I've heard that you can use 256x256 videos, and that as long as the videos are clear, it's perfect for teaching motion. I've also heard that if you accompany this with some high resolution images of the same subject, it can greatly improve the results. Clear motion, and great detail. That's how the person who did the pussy lora, which has zoom-ins on clear close-ups, did their training.
Thanks for that info. I'll look at it, and see how I can approach things.
@Jellai Yes that's correct. You potentially have to go even a bit lower than that for the videos depending on how long they are, to fit them in 24GB. By the way if you're trying to keep costs to a minimum then vast.ai can be cheaper than runpod.
@xto Do you happen to know if 2x4090s can be trained in the same way as a GPU with 48GB of VRAM? Or does it not work that way? I'm thinking that even if it does work, it would add enough time to the training that I'd be better off renting a single GPU with more VRAM for more money.
I paid, but I never ask for buzz and I think it helps the community when we can make these free, especially in light of civitai's recent censorship policies.
One great thing about the 3D printing community is that virtually nobody asks for money ever, and as a result, we all benefit.
Someone might make a nut and bolt you needed for a fitting, and you release that for free, and someone else uses both of those to make a higher-level design that you might use in the future.
That isn't always the case in AI, but I think we should strive for it.
I train locally even though it takes longer and can be much more complicated to optimize.
Not everyone has a good GPU, but yeah if you have the choice, push for more accessibility.
Did you mean to have two "T's" in "glass wall with reflect lightt"?
Nah that was typo, correct one was 'glass wall with reflect light'
This is awesome! I'm guessing I should have started with normal breasts and then pressed them against the glass in the prompt, or dose t2v not work with i2v generations?