Since I saw that the background was noisy in the previous generation, and I guessed that it could be due to the effect of moore's codebase, I guessed that you still used the "scaled linear" schedule during training, but the inference of moore's codebase is "linear" by default schedule, which is the same as your current inference, so I changed your inference to a "scaled linear" schedule (after aligning it with the training), the background noise was eliminated.
XeoOuYang
commented
3 months ago
Since I saw that the background was noisy in the previous generation, and I guessed that it could be due to the effect of moore's codebase, I guessed that you still used the "scaled linear" schedule during training, but the inference of moore's codebase is "linear" by default schedule, which is the same as your current inference, so I changed your inference to a "scaled linear" schedule (after aligning it with the training), the background noise was eliminated.