lora
#5
by ludashi666 - opened
Is this a model with LoRA integrated into it, or is it a fully fine-tuned model? Also, will a LoRA version be released?
It’s quantized FP8 with some layers kept in higher precision. against pruned FP8 in full-action scenarios, it generate better quality.
It’s a 33b model, training a distilled LoRA for it would take months, or you'd need a multi-GPU cluster. they might release a distilled LoRA or a Lightx2V-style variant
but right now it’s faster than WanN 2.2
from my tests if there isn’t much action and motion like talking-head shots, the 'heun' sampler was the fastest with 3 steps. But when the character is running, you need to increase the shift to 10 and steps to 8 as a minimum.