FLUX image caching delivers 1.125x speedup but struggles with dense interior scenes

A developer benchmarked training-free block-residual caching for 4-bit FLUX.1-dev on an Apple M5 Max, achieving a verified 1.125x speed improvement using a warmup-split caching policy. The approach performed well on portraits, product renders, and typography, with PSNR scores above 29, but dense interior scenes remained a persistent weak point, scoring only 22.62 PSNR. An attempt to improve quality via residual-volatility-gated reuse recovered minimal quality gains of roughly 0.2 dB against a target of 6 dB while also reducing the speedup to 1.093x. The gate was confirmed to be firing but did not trigger more frequently on difficult content classes than on easier ones, suggesting the failure mode is not easily addressed by this mechanism. The author concludes that block caching has reached its practical limit for this use case and frames the findings as a foundation for future investigation rather than a solved problem.