would it work well with 2 pass flow using h3 latent upscaler?
nfe 8 means taking 8 steps in single pass, right?
how can i divide the sigma into pass#1 and pass#2?
is it okay to use as below? which is exactly same when using lightx2v turbo lora.
pass#1: set nfe to 8, connect splitsigma node, set step to 4 and take high noise only
pass2: use manual sigma e.g. 0.9035, 0.8000, 0.6316, 0.3158, 0.0000
Yes 8 steps in a single pass, I actually have a node for sigma splits that I’m testing, I will push it soon. Right now the method you outlined should work, SplitSigmas node and manually split at the Euler values.
Yes 8 steps in a single pass, I actually have a node for sigma splits that I’m testing, I will push it soon. Right now the method you outlined should work, SplitSigmas node and manually split at the Euler values.
thank you for quick response.
so, at pass#2, if i use manual sigma, step counts must be either 4,6,8 in euler values, correct?
i also found the acc scheduler node. is it already okay to use this in pass#2?
It has to be 4 or 8, I added 6 but PDD is rigid and you don't get a result. Yes you can use the node, I just pushed the split scheduler to the main branch.
It has to be 4 or 8, I added 6 but PDD is rigid and you don't get a result. Yes you can use the node, I just pushed the split scheduler to the main branch.
thanks! after adopting it, i tried the first run but it seems to not support the hybrid model?
the model i use for ref2va: minimax_h3_hybrid_fl2va_ref2va_b15-49-int8
[WARNING] MiniMaxH3PDDAccApply: model's adaln table is the fl2va trunk but the PDD file is minimax_h3_ref2va_pdd_acc_8step_comfyui.safetensors — trunk mismatch? Pair FL2VA with fl2va, Ref2VA with ref2va.
the task didn't abort, but it seems to be processed without PDD acc.
I believe PDD is still being applied, but that model has some issues (I've tried it). Those specific hybrids ship int8 codes in a foreign layout with no comfy_quant metadata, so ComfyUI loads the raw quantized codes as weights without their scales. That model is has some internal issues in ComfyUI regardless of PDD. Regardless of the warning, PDD should be applied.
I believe PDD is still being applied, but that model has some issues (I've tried it). Those specific hybrids ship int8 codes in a foreign layout with no comfy_quant metadata, so ComfyUI loads the raw quantized codes as weights without their scales. That model is has some internal issues in ComfyUI regardless of PDD. Regardless of the warning, PDD should be applied.
Hmm, I observed the severe ghosting issue tho. might as well going back to pure ref2va.
thank you for the effort!