Frequency-filtered noise in the diffusion forward process steers what the denoiser learns, yielding modest FID gains on some datasets and partial recovery after known-band corruption.
Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Shaping Inductive Bias in Diffusion Models through Frequency-Based Noise Control
Frequency-filtered noise in the diffusion forward process steers what the denoiser learns, yielding modest FID gains on some datasets and partial recovery after known-band corruption.