For some reason I'm not seeing Wan character loras. Is there a reason? Tell me if there is. I think I may have been first, but the flood gates appeared to have opened. Great to see mor Wan character loras! Wan v1.0 is trained on 14b at 480p. Samples are upscaled 2x.
Hunyan V1 was trained at 512x512, V2 used the same dataset, but trained at 768x768. Samples for Hunyuan V2 are 448x720, and not upscaled. Same prompts.
Shannon Louise Eaves born 1966, active 1984 to 1990 under the name Nikki Charm. According to IMDb she was 4ft 11 inches, and Blonde, adorable, enthusiastic and cute as a button, she always had a big grin.
Lora trained using ttplanet's Windows GUI for Kohya's musubi-tuner. Trained mostly from screencaps of very low resolution videos from the 80s. 72 images, 120 epochs
Samples produced with hunyuan_video_FastVideo_720_fp8_e4m3fn, 400x640, 8 steps, for 121 frames. 4x vti multiplier using Rife Tensort and rife49. Upscaled 2x with Topaz Video.
Prompts used:
nikkic, closeup video of a young woman with platinum blonde hair in a ponytail, on the sidelines of a football game. She is wearing a red and white cheerleader uniform.
nikkic, closeup video of a young woman with platinum blonde hair, standing on the pitcher's mound, holding a microphone and singing the national anthem at a baseball game.
nikkic, closeup video of a young woman with fluffy platinum blonde hair, walking slowly towards the viewer on a path through a park on a summer day,looking from side to side. she is wearing business casual
nikkic, closeup video of a young woman with fluffy platinum blonde hair, having lunch in a coffee shop. she is wearing business casual
nikkic, closeup video of a young woman with platinum blonde hair in a ponytail, having lunch in a coffee shop. she is wearing business casual
Description
Retrained at higher res
FAQ
Comments (4)
I train lora 4070ti 12Gb, 71 pictures, 500px*500px, 100epochs, and the display takes 36 hours, is this correct?
Assuming no repeats, I'd guess that's probably about right. my 3090 would probably do that in about 20 - 25 hours.
That being said, and I haven't tried WAN training yet, but I've trained about 30 LoRAs for Hunyuan, and with about 30 images, I usually find that 20 epochs at 5 repeats (using Diffusion-Pipe) is plenty to get a great likeness.
@kmdcomp I think that is about right for Hunyuan, and it could be that people give up at that point with Wan. There are 48 images in my dataset, 250 epochs, no repeats. I first did 50, then resumed with another hundred, and then 100 more. Overall about 7 hours with a 4090.
@jpXerxes That might be why people are saying that training on WAN is harder than Hunyuan. With so many variables though, it's hard to rule out errors in settings, images/captions, and so on. I plan to try WAN training overnight and see what I get.