DHM · fotello raw-test · 2026-08-11

Every pipeline against the editor, on two never-seen shots

Two ambient AEB interiors, no flash on any frame, take 1 only (3 brackets each) so every pipeline sees identical input. Now scored against the real editor beat that shipped for this property (Adobe Camera Raw / Photoshop, 2026-04-08) rather than against the base. Everything is normalised to the same 3:2 / 2496×1664 geometry; the editor's crop aligns with ours at zero shift, so one window mask covers all. Fotello is a competitor AI service — the file is a screenshot of their app, watermark and UI bar included; the UI sits in the bottom ~11% and does not touch the window mask.

Two results are stable across both shots; the rest of the ranking is not. Stable: the v4-caption arms fail on both (36% / 32% and 55% / 40%), and they are the only arms that fail. Stable: Golden-A holds on both (124% / 121% shipped, 124% / 134% no-grade), as do Fotello (106% / 103%) and the base itself (107% / 104%). Unstable: every Golden-B lift arm swings hard between the two rooms — v3 lift@1.0 is 153% on the living room and 57% on the dining room. With n=2 the mean column hides that, so read the per-shot columns and the spread. The one comparison that survives on both shots is the controlled pair: v3 lift@1.0 (153% / 57%) versus v3 v4@1.0 (36% / 32%) — same LoRA, same scale, same base, same seed, prompt the only difference, lift ahead in both rooms.

Read luminance alongside detail. The editor sits at L 210.7. Golden-A shipped (208.6) and the base (207.5) sit right on it; Fotello goes considerably darker (179.6) with almost no clipping, which is an aggressive window-recovery look; the shipped v3 sits at 239.3, far brighter than the editor, which is the washed-out window you can see in the crops.

Every Golden-B arm appears twice: the raw fal edit, then the same edit with its matched Stage-3 LUT (fitted_curves_b2 for v1/v2, fitted_curves_v3ablation for v3). The +LUT cell is the full shipped chain (base → edit → LUT); the raw cell isolates what the generative stage alone did. Golden-A has no LUT anywhere in its path — its tone work is deterministic ops (whitepoint lift, exterior grade) and the develop is the deliverable. Note these LUTs are calibrated to the Scotch Pine editor, not this property's, so they are a tonal approximation here.

Caveats. Two shots, windows only. Detail above 100% is not automatically better — it can mean more local contrast than the editor chose. Golden-A ran GP-a3cfdd43, not the v7.2 chain (step1x-dhm is no longer deployed). Golden-B's Stage-0 skipped the VM window composite (no VM sidecar; SPEC allows it).