Give the score a specific job
VirtualStagingFlow’s Room 08 record has an appealing technical result. Framing passed, and architecture drift is 9.7. The note calls it the lowest drift among the tool’s tested rooms.
The same note says the kitchen was barely staged. Only accessories were added.
That makes this a useful example for anyone building a review interface around benchmark data. The automatic result and the task completion question need separate places in the interface. For this output, I would display the structure measurement beside the recorded description of what changed.
The VirtualStagingFlow benchmark record concerns real CC0 apartment photographs. The broader benchmark remains partial. Blind human scoring is planned, so I would not present these records as a completed blind comparison.
Read the kitchen record in context
Room 08 is the loft kitchen photo. Its recorded settings were Stage a photo, Contemporary, Stage for General, and Clear first. The room type was Kitchen, with no custom request.
The source image was 1600x1200. The output was 2048x1536, and its aspect ratio was preserved. Alignment succeeded. The score file reports 173 inlier matches.
Those facts help describe the technical result. They do not change the written observation that the staging amounted to accessories.
For a review application, I would show a task completion prompt beside this row: is the amount of furnishing appropriate for the intended kitchen image? The record does not establish a universal amount of staging that every kitchen needs. It gives a concrete reason to inspect this particular result.
I would also keep the settings visible during that inspection. A reviewer should see what the operator requested before deciding whether the delivered change was sufficient.
A different reason to pause
Room 09 shows why an exception queue belongs ahead of any numeric ranking. Its architecture drift is 5.7, but alignment failed. Framing is unknown.
The record explains the failure. The original’s nearly blank white walls supplied only 5 keypoints in the evaluated top 65% of the frame. The method needed 25 inlier matches. The score fell back to an unaligned measurement, with no long straight lines available to check.
The note explicitly says to read this as unmeasured rather than near perfect. I would route that row to manual inspection before allowing its drift value into a ranked view.
This is a different review decision from Room 08. The kitchen has a successful measurement and a question about the amount of staging. Room 09 lacks a reliable aligned measurement. Keeping those reasons visible helps a reviewer decide what to inspect next.
Build a small exception queue
The downloadable results data has a top level tools array. A tool’s initial room record is stored in room01. Additional records, where present, are in rooms, with a room field identifying each one.
The following snippet reads that structure. It flags explicit alignment failures and unknown values in selected review fields.
import json
from pathlib import Path
data = json.loads(Path("data/results.json").read_text())
checks = ("framing_pass", "architecture_drift", "aspect_ratio_kept")
for tool in data["tools"]:
rooms = []
if isinstance(tool.get("room01"), dict):
rooms.append(("01", tool["room01"]))
rooms.extend((r["room"], r) for r in tool.get("rooms", []))
for room_id, result in rooms:
reasons = []
if result.get("alignment_ok") is False:
reasons.append("alignment failed")
for key in checks:
value = result.get(key)
if value is None or (
isinstance(value, str) and "unknown" in value.lower()
):
reasons.append(f"{key} unknown")
if reasons:
print(tool["id"], room_id, "; ".join(reasons))
The identity check against False is deliberate. Some room01 records omit alignment_ok. The snippet does not turn that omission into an explicit failure. For the selected review fields, missing values, null values and strings containing unknown trigger review. A known false framing result remains a known failure, rather than being relabelled unknown.
This is a selected field check, not a recursive search through every null in the file. For example, an empty custom request belongs to the recorded settings and does not trigger this queue.
Keep completion review beside technical review
Running the snippet on the supplied data flags several rooms across different tools, including VirtualStagingFlow Room 09 for failed alignment and unknown framing. Room 08 does not enter that queue.
That is the intended limit of the code. It finds measurement exceptions. It does not decide whether a kitchen has been furnished enough.
The written methodology describes drift as a proxy and says it should never be used alone. Room 08 gives that caution a practical purpose. After checking whether a score is usable, return to the recorded task and the delivered image. For this kitchen, the next review question is about the work completed.
For Room 08, I would keep a completion review open even though the exception queue is clear. The reviewer could approve the limited accessory changes if they match the intended brief, or request a more substantial result. The important implementation choice is to make that decision explicit. An empty exception list should never silently become approval of the image.
Top comments (1)
I'd keep a small table-driven test beside that snippet: architecture_drift of 0, framing_pass of False, a missing field, and the string unknown. Each should retain its distinct meaning. That would catch a later refactor to a generic 'if not value' check, which would incorrectly turn a valid zero drift score into an unknown measurement.