If someone care to try RVM i get the best results with these parameters:
1.)using rvm_resnet50.pth model instead the dafault rvm_mobilenetv3.pth - i get too many transparent spots with mobilenetv3 in composition output
2.)set downsample_ratio 0.1 if its higher it get more transparent spots and its slower
3.)seq_chunk 4 using about 12GB VRAM with 8K vid , slightly faster then seq_chunk 1 - about 1.9s/it
in summary i can say RVM works good but unfortunatelly its slow . It takes about 150 mins for 90 seconds 8K vid on my 4060Ti 16GB my CPU is very old indeed . I guess the models are too old and are not good optimized for todays hardware and HQ VR vids .
