Common Diffusion Noise Schedules and Sample Steps are Flawed · HackerPapers discussions

Common Diffusion Noise Schedules and Sample Steps are Flawed — mentioned by GaggiX

Wed, 30 Aug 2023 13:04:54 GMT

It has to be said that the use of CFG is not free, the difference between training and inference change the sample distribution. This difference makes it easy for a trained eye to distinguish a SD/MJ image from a real one just from a thumbnail, compare to the text generated by ChatGPT that is pretty much indistinguishable given that there is no discrepancy between training and inference. SD had another flaw that would make the sample distribution different than the one it was trained one but it's easily fixable with new models: https://arxiv.org/abs/2305.08891, aka the model was never trained…

Paper and discussion history. A mention is not an endorsement.

Common Diffusion Noise Schedules and Sample Steps are Flawed — HN submission

Thu, 18 May 2023 00:43:06 GMT

Common Diffusion Noise Schedules and Sample Steps Are Flawed

9 points; 1 discussion reply on HN. Paper and source history. A submission is not an endorsement.

Common Diffusion Noise Schedules and Sample Steps are Flawed — HN submission

Wed, 17 May 2023 03:56:08 GMT

Common Diffusion Noise Schedules and Sample Steps Are Flawed

3 points; 0 discussion replies on HN. Paper and source history. A submission is not an endorsement.