Perceptibility of the fixed-filter trade-off

Determine whether the reduced arousal-control response of the image-agnostic fixed filter, which incurs no inference cost, is perceptible to users and consequently whether it is a preferable deployment target to the full image-conditioned model.

Background

The authors distill the learned model into an image-agnostic fixed filter by averaging predicted transformation parameters for each conditioning value. This fixed filter has no inference cost because it runs entirely in the shader, but its arousal response is 15.3% shallower than that of the full image-conditioned model. The paper does not test whether users can perceive this difference, leaving the practical choice between the computationally cheaper fixed filter and the stronger full model unresolved.