This lora has been trained with over 5000 clips, improving human movement and anatomy. It can also interact with non-human objects. No trigger word is required.
Training process: Videos were scraped from the internet and trimmed to between 5 and 15 seconds. Tagging was done by hiring freelancers; 30% of the dataset was tagged and other 70% tagged by llm, and GPUs were rented at Runpod. Future versions will have a higher percentage of the dataset tagged and more training time.
Description
FAQ
Comments (13)
5000 clips mean nothing, show me how your captioning your data set
Based on the samples, there's no way it was good.
Can you recommand a workflow to make this LoRa work ? I can't make it not generated slug
Works great for me in Wan2GP
Bro what are these samples? Were these captioned by aliens that never seen porn?
lot of room for improvement
man thank you for sharing but...
if you describe all the work you did, even hiring freelancers, and then you only post 2 meh demo vids, people are going to think this is fake, don't you think?
The examples are not the best BUT the lora works actually really well! Sounds like many that have commented here have not actually tried it. Thank you and keep it up :)
I just don't have the space or time to try out models that only have garbage samples. No one has posted any examples of this lora other than the other, so i'll stay away.
Would love to see some samples you made !
I don't know who had trouble getting it to work or who dislikes the sample images... Everything works fine for me; the anatomy has improved, and the movement looks better too. That said, there is something to criticize the author for: this LoRA seems a bit too aggressive regarding the character's appearance—it has a tendency to try and "tweak" things every now and then.