The purpose of this lora is to increase the clarity and sharpness of Hunyuan videos.

One of my main frustrations with Hunyan/WAN etc video has been how grainy or fuzzy the results are unless you render them at resolutions that are out of reach for those of us mere mortals with retail level video cards. So I set out to see if I could fix that to some degree.

This lora was trained on 6 different 60fps videos ranging from 10s to 22s in length. 4 of the videos were pornographic and 2 were 60fps demo videos of movement involving inanimate objects. As a result it seems to work best on human beings and respectably well on everything else. It also seems to help in creating a better likeness when combined with loras meant to depict a particular person.

The lora was created using diffusion-pipe (b5ea041..13a96a5) on an NVIDIA 3090FE. It was run for 30 epochs (450 global steps) with repeats set at 10.

Recommended strength: .85 to 1.