lightx2v

/

Wan2.2-Lightning

like 221

I don't think it's LightX2V's fault, I think it's the fact that WAN 2.2 was split into two models, which has always been a point of failure in the AI space. Just like with SDXL when they required a refiner model, no one wanted it, no one used it, and it created tons of extra overhead. The leap from WAN 2.1 to WAN 2.2 isn't really all that much, and not really worth it for double the model size and double the processing time. WAN 2.1 is still much more stable.

I think part of the problem could be that WAN 2.2 stuff is LoRA's now rather than a checkpoint like with 2.1. The full checkpoint baked into WAN 2.1 delivers generally better output than using the LoRA instead. Personally I will probably skip over WAN 2.2 and hope WAN 2.3 goes back to a single checkpoint. The high noise model is really just a crutch to try to force more movement by creating more noise. As we've seen with previous image generator architectures, more training on a single model can deliver vast increases in capability, and there are also a lot of other ways that have made WAN 2.1 produce better results, including custom samplers, fiddling with Shift settings, etc. So I think there were just better ways of getting more out of a single model, instead of splitting it into 2 models. WAN 2.2 Low Noise is close to WAN 2.1 single model, but they crippled it to make it play well with the high noise model, which is clearly why the low noise model can generate stuff on its own, but isn't too great when used on its own, even if it can in some cases give really good results, it requires too much fiddling and setting changes per each image/prompt used in order to make it coherent.