Not sure exactly where the issue lies, but in the process of fiddling around I found that scaling the output samples by 2x before the decoder when using single-step generation fixes that instance of it. The above images are from sample_t2i_dpgbench_ddp.py
Not sure exactly where the issue lies, but in the process of fiddling around I found that scaling the output samples by 2x before the decoder when using single-step generation fixes that instance of it. The above images are from
sample_t2i_dpgbench_ddp.py