Two packs applied via ffmpeg lut3d to the v1+lift scale-1.6 output:
purple11 (12 cubes, ~/DHM/LUTS on sandstorm) and highway (15 cubes,
NAS: Voltage Labs/clients/DHM/real-estate-luts-fcpx-apple-motion….zip).
Sorted best→worst by mean |ΔL| at p25/p50/p75 vs beat (beat: 45.9 / 69.9 / 75.7).
Baseline RAW sits mid-pack; only smooth-sailing, hw-11, urban-cowboy, hw-06 beat it. Two fitted cards = percentile curves calibrated on the other 24 holdout scenes (s009 unseen). Score is tonality-only — check color casts by eye.
RAW→base prototypes, develop→align→enfuse from the 6 brackets. A = plain (non-guide reproduction).
B = tuned: exposure-optimum 0.42 + VM-bbox window composite from darkest bracket + neutral WB. B has 0.00% clipped pixels and real exterior detail in every window.
base A · plain enfusebase B · tuned (VM window composite + WB)old non-guide input (what v1 LoRA saw)base B → fal (v1+lift s1.6, seed 42)base B → fal → fitted LUT (full new track)
Hi-res test · s009 @ 2560×1728 (2026-08-07)
FLUX.2 accepted 2560×1728 natively (19s). Edge damage vs base nearly halved (0.020 vs 0.036) but tone drifted (+1L vs 1536 track). LUT needs a hi-res refit before adopting.