EDIT: It has come to my attention that the musubi-tuner version I used for this had a bug that affects the quality of the LoRA produced, so I will be re-training this one soon for a v2.
This is my first HunYuan LoRA. No, really, it's my very first one. I think it turned out nicely. This is the final save from the run and it seems great, but it's possible it could be overcooked after 4000 steps.
I used frames selected from clips trimmed from season one episodes of Bewitched, which aired in 1964, so the LoRA has that old television look to it.
I used the GUI musubi trainer from this civitAI user here:
https://civarchive.com/articles/10335/hunyuan-video-lora-trainning-with-gui-in-windows
I'm not gonna lie this shit kicked my ass and I wouldn't have gotten it working without the comment section and lots of GPT-1o help. It was difficult as hell for me.
Anyway, looking for feedback. Please share clips or comment.
Description
Retrained on frames from the 1964 season of Bewitched and portrait photos/face crops from the 60s using musubi-tuner.
Sooo much more versatile and crisp.
If the model seems stiff, drop the strength in increments of .05 and reassess. This one in particular seems good at 1, however.
If you have any feedback, please share. As this is one of my first projects with HunYuan, I ended up having to train numerous versions of this LoRA during my testing and learning experience.
So yeah, postive or negative, I'd love to hear it.