Training method ？
wonderful lora！ May I ask what is your training framework？
（1） is it single-stage or multi-stage training ？
（2）do you use DMD or other distillation methods like CM or rCM？
I myself train with DMD but find the video gets over-saturation under 4 NFE
hi, I am also training 4 step lora using DMD2 and find that the video gets over-saturation after 200steps student update.
hi, I am also training 4 step lora using DMD2 and find that the video gets over-saturation after 200steps student update.
so do I, so I guess there maybe some other distillation techniques beyond dmd2 used in larryvrh's training to avoid over-saturation
hi, I am also training 4 step lora using DMD2 and find that the video gets over-saturation after 200steps student update.
so do I, so I guess there maybe some other distillation techniques beyond dmd2 used in larryvrh's training to avoid over-saturation
Did you tried other methods like CM or rCM? I am now training CDM and dual lora like z-image++ turbo, but not work maybe. And I find that Lightx2v's DMD2 training set(like LR) helps to reduce motion blur.
hi, I am also training 4 step lora using DMD2 and find that the video gets over-saturation after 200steps student update.
so do I, so I guess there maybe some other distillation techniques beyond dmd2 used in larryvrh's training to avoid over-saturation
Did you tried other methods like CM or rCM? I am now training CDM and dual lora like z-image++ turbo, but not work maybe. And I find that Lightx2v's DMD2 training set(like LR) helps to reduce motion blur.
I did not meet motion blur issues. I tried Lightx2v's DMD2 training code and my own dmd2 training code, they looks good even under 4nfe without blur or artifacts. But the color seems not similar to multi-step baseline, for example, tend to be warmer than baseline(but not obvious over-saturation). I have not try rCM but I guess that maybe helpful.