This lora was trained with artificialguybr, with only one video, this.
Description
FAQ
Comments (14)
That's amazing, and only with one video!. Do you think you will be able to make one with robotic motion? Would be great to see.
I would love to hear anything about the process wrt training this LoRA for high and low with a single short clip. Also, is "hard" like "strong"? "Soft" is a weaker effect and "hard" is more effective at the same strength?
I think she wrote a guide already?
@LDWorksDavid - would love to see it. Just checked their civit, HF, twitter, and rentry pages and see nothing related to this or Wan.
Yeah that's crazy, please share the technique!
i wrote an article related to qwen, the process is quite the same, but no wavelet and sigmoid on this config, just standard one
@luisa_pinguin - the process of training a qwen LoRA on images... is not the same at all as training a video LoRA on a single video. My question isn't "how do I train a LoRA?" it's ""how did you train a video LoRA on just one video?"
@Gongoloid so, im my article, i train a lora with few datasets, with 4 dim, 1 alpha with the minimal config possible, in the video, i did the same, but with one video, with 150 steps. the soft is 100 steps, the hard is 150 steps.
keep 4 dim, 1 alpha for "glue" on video, this also helps the video learn faster.
@luisa_pinguin - I have trained hundreds of video LoRAs, I am only concerned with this specific case - I don't need to learn how to train LoRAs.
So your dataset was just one video and you didn't do anything else differently with the data? That is the whole question. I appreciate your response and thank you for the LoRA and information.
I use musubi-tuner for Wan and I think I'll just choose a distinctive video and try a single video LoRA sometime over the weekend.
Thanks for the inspiration.
@Gongoloid no, the data is one video, nothing else.
now im testing 1 second video.
@luisa_pinguin How it went?
@LDWorksDavid is the jumpscare i2v
@luisa_pinguin Thanks!