best sampling settings for the 4 step ref2v?
is it me or is it not working / not good?
I tried it in comfy and sa solver ( er_sde/ and euler / with beta or beta 57 didnt work at 4 steps what settings, thanks
didnt work at 4 steps
None of the existing or future 4-step LoRAs work properly at 4 steps. You need at least 6-8 steps.
yeap lowest i reach is 6steps but 8steps is the ideal to retain details! no lora seems to work at 4 stepst!
Is the best weight 0.75 or 1.0?
I tried strength 1, It is good. I found that it has sound problem (for audio shift 4), (shift 6 is good), can causes cross some language merge like Cantonese combine with Mandarin.
euler linear quadratic works for me, just use more resolution and maybe 6 steps
I tried strength 1, It is good. I found that it has sound problem (for audio shift 4), (shift 6 is good), can causes cross some language merge like Cantonese combine with Mandarin.
Thanks for sharing . I would like to know how to set the shift value. Which node should be used for that ?
Distillation requires ODE samplers, specific sigmas and correct shift values. If using Euler/simple produces garbage it's because the model is bad.