Parameters without lightx2v ?

[Emen0s]

Hello,

The model is working perfectly fine with lightx2v on.
But I want to use the full model without lightx2v to get better quality (lightx2v removes some fine details and film grain that I want to keep).

But when I try to remove the lightx2v LORA, I get some distorted hands / morphed hands on static / small movement.

I tried changing the parameters this way :

Shift : 5 (too high ?)
Steps : 30-40
CFG : 5
Resolution (portrait) : 704x1280

I tried adding the DPO Lora, I think it actually helped a little.
I tried using the DWpose as input too.
Using comfyui.

Is there any recommended parmeters to have a good results without the lightx2v ?
Maybe I need to adjust the prompt ?

Thank you for you work and your time !
Have a nice day !

[fengjia-guo]

Hi, thanks for this feedback.

Our default settings without LightX2V are Shift: 5, Steps: 40 - 50, and CFG: 4. In one of my earlier experiments, I found CFG = 4 better than CFG = 5, but, I'm afraid that tuning sampling parameters may not help much in distorted hands. LightX2V reduces the error accumulation during sampling, so completely removing it (then you have to use more steps) can lead to distorted body or worse motion quality.

If you're concerned about the details, you can try reducing the strength of LightX2V or increasing sampling steps at the same time (e.g. from 4 to 8). You may also refer to the experiences from other users in texture, e.g. #23 and #20.

We're working on the release of training code and believe that the issue of details can be better resolved with some LoRAs of SCAIL-2 model, or maybe a distillation LoRA trained for SCAIL-2.