A benchmark's photo credits and its measurements describe the same inputs from different directions. The credits record where a photograph came from and its original and saved dimensions. The measurements record the input dimensions associated with each staged result.
Comparing them is a useful consistency check. If a photograph's saved dimensions appear nowhere among the measured input dimensions for its room, something needs explaining. An input may have been resized, a measurement may refer to an older copy, or the room identifier may have been assigned incorrectly.
StagingBench publishes enough information to make this comparison directly. Its photo names begin with a room identifier. For example, 01-open-plan-kitchen-living.jpg belongs to room 01. Keeping that identifier as text preserves its leading zero.
The following Python snippet uses only the standard library. Run it in a directory containing credits.json and scores.csv. It counts license labels, gathers the input dimensions seen for each room and reports photographs whose original and saved dimensions differ.
import csv import json from collections import Counter, defaultdict with open("credits.json", encoding="utf-8") as handle: photos = json.load(handle) sizes_by_room = defaultdict(set) with open("scores.csv", newline="", encoding="utf-8") as handle: for row in csv.DictReader(handle): sizes_by_room[row["room"]].add( (int(row["orig_w"]), int(row["orig_h"])) ) licenses = Counter(photo["license"] for photo in photos) mismatches = [] downscaled = [] for photo in photos: room = photo["file"][:2] original = tuple(photo["original_size"]) saved = tuple(photo["saved_size"]) if saved not in sizes_by_room[room]: mismatches.append(room) if original != saved: downscaled.append((photo["file"], original, saved)) print("License counts:", dict(licenses)) print("Rooms with no matching saved size:", sorted(set(mismatches))) print("Photos with changed dimensions:", len(downscaled)) for filename, original, saved in downscaled: print(filename, original, "->", saved) The set for each room matters. A room appears in multiple measurement rows, and those rows should not overwrite one another during collection. Storing all observed dimension pairs lets the comparison ask the intended question: does the credited saved size match any measured input size for this room?
An absent room also produces a mismatch. Its set contains no dimension pairs, so the credited saved size cannot match. That makes a missing room visible without requiring a separate branch in the snippet.
Applied to the published data, the license counts are 9 CC0 and 1 Public domain. All ten saved sizes match a dimension pair recorded for the corresponding room. The mismatch list is empty.
The dimension comparison identifies five changed photographs: the attic, the living room with the bay window and fireplace, the bedroom, the living room with the large window, and the small white room with double doors. In each case both dimensions decreased, so all five were downscaled. The other five retained their original dimensions.
The snippet uses inequality to select changed dimensions. In a different collection, that condition could also identify enlargement or a change in proportions. Calling every such result downscaling would require checking the direction of the change. Here, the recorded widths and heights both decreased for every selected photograph.
This check helps catch silent resizing because it connects the declared saved image to the measured input. Suppose an input is replaced with a smaller copy while its credits retain the previous dimensions. If the measurements reflect the replacement, the comparison exposes the disagreement.
Stale measurements can produce the same symptom in reverse. The credits may describe the current input while the scores still describe an earlier version. The script cannot decide which record is current, but it gives a concrete room identifier to investigate.
There is also a limit to the matching rule. It accepts a saved size when any observed pair matches. If a room has several different measured sizes, a match can coexist with inconsistent rows. A stricter audit could additionally report rooms with multiple input dimension pairs.
Matching dimensions also cannot prove matching image content. Different photographs can share a width and height, and edits can preserve both. This is a narrow check of declared sizes and identifiers, not a complete reproduction audit.
The photo credits and published results provide useful starting points for scrutiny. Checking their agreement makes the input preparation easier to understand before interpreting the measurements built on it.

