Could you explain the differences between “Wan2.2-Distill-Loras” and “Wan2.2-Lightning-Loras”?
Hi, similar issues can be found here https://github.com/ModelTC/Wan2.2-Lightning/issues/31.
In summary, loras in https://huggingface.co/lightx2v/Wan2.2-Distill-Loras comes from SVD while loras in https://huggingface.co/lightx2v/Wan2.2-Lightning comes from lora training.
Hi, similar issues can be found here https://github.com/ModelTC/Wan2.2-Lightning/issues/31.
In summary, loras in https://huggingface.co/lightx2v/Wan2.2-Distill-Loras comes from SVD while loras in https://huggingface.co/lightx2v/Wan2.2-Lightning comes from lora training.
Which is better: the distill LoRA obtained by applying singular value decomposition (SVD) to the full parameter differences, or the Lightning LoRA trained with standard LoRA configurations?
Newer are better, i.e distill loras and models (1030 for the high) are better then the lightning loras
Newer are better, i.e distill loras and models (1030 for the high) are better then the lightning loras
Thank you!