Open sdbds opened 1 year ago
I've been experimenting with a model trained on v-prediction with zero terminal SNR, including with K-diffusion samplers. I've also applied the modification to the DDIM sampler under LDM and tested it out and it works as described -- being able to produce a solid black image with a dark subject. Not applying the zero terminal SNR beta rescale does still produce results close to this -- far closer than anything achieved while the model was being trained on epsilon loss with zero terminal SNR -- but does not quite achieve the solid black background. So as far as what would need to be done on K-diffusion's end to bring it up to this capability, including information on my experiences experimenting with these so it can be known what these should be expected to do (going off the paper's suggestions in order):
I would love to see K-diffusion support this fully. Zero terminal SNR models are incredibly capable.
https://arxiv.org/abs/2305.08891
I think these might be helpful