Deep learning image filters lack normalization. We propose filter normalization with learnable scaling and shifting, achieving significant improvements in intensity variation benchmarks.