2-stage + latent upscale case.
Hi,
What would be best usage of veda in this case:
hybrid model, pdmd speed lora with 4 steps, ref2v node, refmods.
(refmods like fl or hybrid model from what I tested)
1st sampler around 0.6mpx,
then latent upscale to 1mpx or higher,
then 2nd sampler 4 steps (separate model path without speed lora, this lora burns result as i tested) with low denoise.
So, how do you think? Should I put veda node before both samplers, or just 2nd one (where speed gain shoud be bigger)?
Btw, I tested with single stage generation, speed gain is like this on my 4060Ti / 1mpx:
4/4 [05:22<00:00, 80.73s/it] - veda
4/4 [08:37<00:00, 129.32s/it] - native
Thanks!
Thanks for sharing your benchmark. We trained on 8step LoRA, but it should be able to adopt other settings including other few-step LoRA and style LoRA. We only test few of them, so if something work or not, we are happy to know!
For Veda, the speed-up scales largely with sequence length (resolution x time), as Veda mostly scales with n (with a tiny n^2 factor) but naive attention scales with n^2. We would suggest put it right before any H3 model node.