Thanks @buoyancy99 for the great work.
A question on the training scheme: with independent per-frame noise, sequences may have one frame at very low noise while others are high. In low-motion content, adjacent frames are nearly identical, so the model can minimize loss by essentially copying the near-clean neighbor rather than learning to denoise without an anchor.
Did you observe this in practice? And if so, did you do anything to mitigate it (loss weighting, min-noise floor, correlated noise across frames)?
Thanks @buoyancy99 for the great work.
A question on the training scheme: with independent per-frame noise, sequences may have one frame at very low noise while others are high. In low-motion content, adjacent frames are nearly identical, so the model can minimize loss by essentially copying the near-clean neighbor rather than learning to denoise without an anchor.
Did you observe this in practice? And if so, did you do anything to mitigate it (loss weighting, min-noise floor, correlated noise across frames)?