What the AI image is allowed to do

The image may propose movable furniture, rug, lamps, small plants, and color relationships. It may help you compare layout ideas and discover what to measure. It may not redefine the room, hide inconvenient services, infer exact dimensions, guarantee furniture fit, select products without specifications, place an order, or justify construction. The retained ceiling is READY_FOR_MEASURE_AND_OWNER_REVIEW.

LayerLocked or allowedEvidence boundary
ArchitectureCamera, door, window, radiator, outlets, pendant, floor, walls, ceiling, baseboardsMust stay visible; image comparison is still not a survey
Movable designSofa, chair, coffee table, rug, floor lamp, small plantsConcept only until measured and specified
ClearancesDoor swing, route, radiator space, outlet accessVisual check creates questions; physical measurement decides
ReleaseOwner review after 11 selected checks and 7 mutation rejectionsNo purchase or construction approval

Before Step 1: decide what the photo must not reveal

A real room photo can expose faces, family images, mail, medication, calendars, addresses, screens, keys, security devices, reflections, window views, or expensive belongings. Remove what the design task does not need. Keep an unredacted original privately, export a sanitized working copy, and review the current service’s retention controls. If the home belongs to someone else, get permission.

Check the service’s retention policy for the account and file workflow you intend to use before uploading a room photograph; this is separate from permission to share the image.

Take the source photo from a stable, useful viewpoint. Include the floor-to-wall boundaries and fixed features you need to preserve. Do not use panorama stitching or a dramatic ultra-wide lens if it distorts proportions. If multiple walls matter, capture separate labelled views rather than asking one image to prove geometry it cannot show.

Exact reader tool path: for the retained run, Codex called OpenAI image generation and editing to build the fictional source room and its variants. No image-model identifier was returned, so that field remains unknown. To reproduce the method in ChatGPT Images, start a new chat, attach room-source-photo.png, paste the furniture-only contract, and request a single layout concept. Download it under a candidate filename. Keep the source open beside it and complete the anchor ledger before asking for another version. If an anchor fails, begin again from the source image and name only the allowed furniture change plus the failed anchor—do not continue editing a structurally drifting candidate.

  1. Attach the sanitized source and record its filename, 1448 × 1086 geometry, and hash.
  2. Paste the immutable-anchor list and furniture-only concept prompt; generate one layout.
  3. Download that image as a candidate, never as a floor plan or “after” photo.
  4. Compare camera, door, window, radiator, outlets, pendant, floor, walls, ceiling, and baseboards one by one.
  5. For a failed anchor, reattach the source, issue the targeted-correction prompt, save a new version, and rerun the full ledger before measuring anything.

Below are direct evidence views of the retained source file and furniture candidate. They do not imitate a product interface; their purpose is to separate the actual image artifacts from the explanatory result boards.

Actual retained file view of the fictional room source photo before AI redesign
Actual retained input artifact — the source view locks camera, door, window, radiator, pendant, outlets, and floor before styling. Credit: Grow AI Skills browser capture of its retained fictional artifact. Open full-size image.
Actual retained file view of the furniture-only room redesign candidate
Actual retained output artifact — the image remains a measurement hypothesis, not a floor plan, fit check, purchase, or construction approval. Credit: Grow AI Skills browser capture of its retained fictional artifact. Open full-size image.

Step 1: sanitize and freeze the source photo

Your action

Inventory private elements, save a private original, create a sanitized working copy, record the crop and pixel dimensions, and define the scope as furniture-only. Our source is a fully fictional 1448 × 1086 room with no people, address, document, screen, key, or location clue.

Room-photo preflight
Confirm the image is sanitized. Freeze the source hash, crop, and camera view. Scope the task to movable furniture and styling. Do not infer an address, dimensions, wiring, or construction plan.

Retained result

The source is classified as a synthetic furniture-only fixture. Its hash and pixel geometry are retained. The fixed features are deliberately clear: one left door, one centered double-sash window, radiator and pipes, two outlet plates on the right wall, a centered pendant, and honey-oak floor.

Step one result board for room photo privacy and source preflight
Step 1 — Remove private details, retain a private original, and declare a furniture-only scope. Credit: Grow AI Skills retained result screenshot. Open full-size image.

Your check

Zoom into mirrors, glossy screens, frames, and windows. A blur or crop must not remove a structural feature you need to lock later.

Step 2: build a protected-structure ledger

Your action

Convert every fixed feature into an explicit no-move constraint. Include count, type, location zone, visibility, and failure action. Then create separate rows for clearances and unknowns. The model should not treat a hidden outlet as proof that it disappeared safely.

Structural anchor map
Lock the exact camera and crop; left six-panel door and brass knob; centered double-sash window and trim; radiator and pipes; two right-wall outlet plates; centered pendant; oak floor; walls, corners, ceiling, and baseboards. Mark dimensions and clearances UNKNOWN.

Retained result

The ledger contains ten rows. C-01 through C-07 cover visual architecture and finish. C-08 through C-10 keep door swing, radiator clearance, outlet position, furniture fit, and budget in measured or unknown states.

Step two result board locking camera door window radiator outlets pendant and floor
Step 2 — Ten constraints separate visible architecture from measurement and purchase unknowns. Credit: Grow AI Skills retained result screenshot. Open full-size image.

Your check

Add vents, returns, detectors, panels, switches, thermostats, sprinklers, balcony doors, egress paths, and rental restrictions visible in your room. If an anchor is not visible, write UNKNOWN; do not assume it is absent.

Step 3: write a furniture-only design contract

Your action

Separate four things: immutable architecture, allowed movable changes, functional requirements, and aesthetic preferences. Our goal is a warm practical Japandi reading room. Allowed changes are a compact sofa, reading chair, coffee table, flatweave rug, floor lamp, and small plants. The door route, radiator airflow, and outlet access remain functional constraints.

Furniture-only contract
Change only movable sofa, chair, table, rug, floor lamp, and small plants. Preserve every locked anchor. Keep door route, radiator front, and outlets visibly clear. No built-ins, drilling, shelves, electrical work, new openings, inferred dimensions, product fit, purchases, or construction.

Retained result

The contract contains six movable item classes, seven immutable visual anchors, three measurement or purchase boundaries, and the READY_FOR_MEASURE_AND_OWNER_REVIEW ceiling.

Step three result board defining a furniture-only room redesign contract
Step 3 — The contract allows movable furnishings but forbids construction, hidden services, inferred dimensions, and purchases. Credit: Grow AI Skills retained result screenshot. Open full-size image.

Your check

If the design goal cannot be achieved without moving a fixed feature, stop the image-only workflow. That is a different project requiring real measurements, permissions, plans, and appropriate expertise.

Step 4: generate a small set of bounded concepts

Your action

Ask for two or three meaningfully different furniture arrangements, not dozens of decorative variants. Repeat the structure lock in every request. Keep the source visible beside the output and label each image as a concept, not the room.

Bounded concept generation
Create a warm Japandi furniture concept from the source. Preserve camera, door, window, radiator, two outlet plates, pendant, oak floor, walls, ceiling, and baseboards. Use movable furniture only. Keep service access and path visible.

Retained result

The run retains three AI edits: a deliberately structural-drift negative control, a visually attractive candidate that still hides outlets and creates door-swing ambiguity, and a targeted correction. Keeping failures is useful because it shows exactly where a broad “redesign this room” instruction becomes unsafe.

Step four result board showing source failed and corrected room redesign concepts
Step 4 — A small concept set exposes both structural drift and obstruction before measurement. Credit: Grow AI Skills retained result screenshot. Open full-size image.

Your check

Do not continue merely because one image matches your taste. First check whether it still depicts your room.

Step 5: fail structural drift and hidden services

Your action

Compare source and concept feature by feature. Fail any changed count, shape, position, size relationship, or visibility. Our controlled bad concept turns the left door into an arch, widens and lowers the window, removes the radiator and outlets, replaces the pendant with recessed lights, and changes oak floor to stone.

Anchor audit
Audit source versus concept. Return ANCHOR_AUDIT_FAIL for any moved, widened, removed, added, covered, or hidden door, window, radiator, outlet, light, floor, wall, ceiling line, baseboard, or camera relationship. List the minimum correction.

Retained result

The polished Japandi image fails all six planted structural changes. A second candidate also fails the usable-layout gate because the right-wall outlets are obscured and storage sits in a questionable door-swing zone. “Furniture only” must include visibility and access, not just unchanged wall pixels.

Step five audit result rejecting changed door window radiator outlets light and floor
Step 5 — The polished Japandi concept fails six structural anchors and cannot be used for shopping or construction. Credit: Grow AI Skills retained result screenshot. Open full-size image.

Your check

Overlay images when possible, but remember perspective and generative texture can defeat pixel alignment. The visual audit is a screening step. The physical room remains the authority.

Step 6: issue a targeted furniture correction

Your action

Return to the source rather than editing a drifting result repeatedly. Name only the failed anchor and the allowed furniture response. The correction removes the problematic storage, adjusts the sofa, keeps the radiator visible, restores two outlet plates to view, and retains the major door, window, pendant, floor, and camera relationships.

Targeted correction
Return to the source. Use a compact sofa, chair, table, rug and lamp. Keep the exact door, window, radiator, pendant, floor and camera. Expose two outlet plates. Keep the door path and radiator front open. Add no construction.

Retained result

The candidate advances as PASS_WITH_MEASURE. Major architecture is visible. Two outlet plates are visible, but exact positions and access still require measurement. The door path and radiator front look open, but pixels cannot prove clearance.

Step six result board showing corrected furniture-only room concept with major architecture visible
Step 6 — The candidate retains major architecture but keeps outlet positions and clearances in the measurement queue. Credit: Grow AI Skills retained result screenshot. Open full-size image.

Your check

Inspect the entire frame after every targeted edit; current image tools can modify outside the named region. If a fixed feature is hidden, return it to unknown rather than assuming it is unchanged behind the furniture.

Step 7: convert the picture into measurements—not dimensions

Your action

Create a worksheet that tells you what to measure in the physical room. Leave the value cells blank. Our eight rows cover wall width, door swing and landing, window sill and radiator, radiator clearance, outlet positions, narrowest walking path, sofa envelope, and total budget including delivery and returns.

Blank measurement board
Create measurement IDs M-01 onward with what to measure, blank value, unit, physical method, and required-before action. Do not estimate from pixels. Include door swing, radiator guidance, outlets, path, sofa envelope, delivery, and budget.

Retained result

All eight value fields remain empty. Door swing is measured by opening the real door; outlet centers are measured from a fixed corner; radiator clearance comes from owner, manufacturer, or qualified guidance; furniture fit follows measured clearances rather than image proportions.

Step seven result board for a blank physical room measurement worksheet
Step 7 — Eight measurement questions remain blank; no dimensions are inferred from pixels. Credit: Grow AI Skills retained result screenshot. Open full-size image.

Your check

Mark furniture footprints with removable tape. Check the delivery route, stairs, elevator, doorway, skirting, plugs, cable bends, cleaning access, and return policy. For tall or tip-prone furniture, follow current product and safety instructions for anchoring.

Step 8: enforce a no-purchase release gate

Your action

Execute the room-specific checker. It opens the source and accepted PNGs and recomputes fixed-crop dHash alarms for the full frame, left door, window, radiator, and pendant; those pixel-derived values must match the run record and remain within fixture thresholds. Its remaining assertions cover four image hashes and equal pixel geometry, ten constraints, six core visual anchor states, the outlet measurement boundary, three clearance or fit unknowns, eight blank measurements, review ceiling, no-purchase and no-construction states, and the six-failure negative control. Seven mutations try to promote the state, claim a purchase, propose construction, invent a dimension, physically replace the window region, weaken the outlet boundary, or alter the candidate hash.

Owner and measurement gate
Run the gate. Reject FINAL, PURCHASED, construction-proposed, invented-dimension, failed-window, unbounded-outlet, and bad-hash states. Return READY_FOR_MEASURE_AND_OWNER_REVIEW at most.

Retained result

All 11 baseline checks pass and all 7 damaging mutations are rejected. Measurements completed: false. Purchase state: NOT_PURCHASED. Construction state: NOT_PROPOSED.

Step eight no-purchase gate with eleven checks seven mutation failures and owner review state
Step 8 — The workflow ends at measure and owner review, with nothing purchased and no construction proposed. Credit: Grow AI Skills retained result screenshot. Open full-size image.

Your check

Complete the worksheet in the physical room, set a budget, choose products from current specifications, and record returns and safety constraints. The image can remain inspiration; the measurements and product documents must control the decision.

Room-redesign failure modes

FailureWhat the image suggestsSafe response
Window is wider or lowerMore light and wall space than existsReject; return to source geometry
Door becomes an archConstruction disappeared into stylingReject; retain exact door and swing
Radiator or vent vanishesFurniture can occupy a service zoneReject; restore feature and obtain real clearance guidance
Outlet is hiddenPower access looks irrelevantMark unknown, reveal it, then measure physically
Camera changesRoom appears larger and furniture fitsReject the comparison; use a fixed source view
Model names a product sizePixels became measurementsLeave value blank and measure the room

Test the structure-lock fixture

Fetch the room audit bundle; its checker reads the PNG lineage, constraint ledger, measurement worksheet, and run record together. Then execute:

mkdir room-redesign-review-kit
unzip room-redesign-review-kit.zip -d room-redesign-review-kit
cd room-redesign-review-kit
python3 -m venv .room-gate-env
.room-gate-env/bin/pip install Pillow
.room-gate-env/bin/python room-redesign-evidence-gate.py

The retained results report 11 selected baseline checks and 7 rejected mutations. These results apply only to the fictional fixture. A real room needs sanitized source photos, its own anchor ledger, physical measurements, current product specifications, owner decisions, and qualified advice where safety or building work is involved.