Kitchen(Lighting) + Bathroom (Taxtiles) one-Variable Test: Prompts, Failures, and Fix Rules

This post is not a “pretty results” showcase. It’s a prompt log + debugging recap from my layout-locked experiments.

I tested two rooms with one rule: Layout Locked + One Variable.

• Kitchen = lighting-only
• Bathroom = textiles-only

The problem I kept hitting is classic: AI preserves the geometry but breaks the meaning/function.

Here’s the full system I used—copy/paste prompts, the 3 failures I saw most, and the 3 fixes that worked best.

My Test Rules (Layout Locked + One Variable)

What “Layout Locked” means in my setup

• Same original photo (reference image)
• Same camera angle + framing
• Same architecture and fixed elements
• No remodeling, no construction, no layout changes

What “One Variable” means

• Kitchen: lighting/exposure/white balance/shadows only
  (No declutter, no decor changes)
• Bathroom: towels + bath mat only
   (No mirror/vanity/tile/light changes)

Copy/Paste: The Lock Sentence (Use in Every Prompt)

I stopped using “no new fixtures” because “fixture” can be interpreted as a light fixture, which becomes ambiguous during lighting tests.

This lock sentence is clearer because it targets architecture and fixed elements directly:

text

keep the exact same layout and camera angle as the input photo, preserve all architecture and fixed elements, same time of day as the original photo, no remodeling, no construction, no layout change, no moving sink, no moving toilet, no new windows, no new doors, no changing tile, no changing flooring

The Prompts I Used in Post 6

Kitchen prompt (Lighting-only)

text
minimal modern bright studio kitchen, lighting-only improvement, brighten the space with clean neutral lighting, reduce shadows, balanced exposure, crisp whites, keep lower sink cabinet exactly the same, keep countertop and sink position unchanged, keep appliances and cabinetry unchanged, no declutter, no decor changes,
keep the exact same layout and camera angle as the input photo, preserve all architecture and fixed elements, same time of day as the original photo, no remodeling, no construction, no layout change, no moving sink, no new windows, no new doors, no changing tile, no changing flooring
--style raw --ar 4:3 --v 6.1 --iw 1.7 --stylize 25 --chaos 0 --seed 22021

Bathroom prompt (Textiles-only)

text
minimal modern bright bathroom, textile-only update, keep the mirror exactly the same, keep vanity, tiles, walls, and lighting exactly the same, swap only textiles: fresh white towels, simple neutral bath mat, clean minimal fabric textures, no decor items added,
keep the exact same layout and camera angle as the input photo, preserve all architecture and fixed elements, same time of day as the original photo, no remodeling, no construction, no layout change, no moving toilet, no new windows, no new doors, no changing tile, no changing flooring
--style raw --ar 4:3 --v 6.1 --iw 1.8 --stylize 25 --chaos 0 --seed 22021

3 Failures I Hit (Layout Matched, Function Drifted)

Failure 1 — Appliance identity drift (Kitchen)

The layout looked close, but the model “re-labeled” objects:

• a built-in refrigerator area started reading like a dishwasher
•  appliances stayed “in the right zone,” but not as the correct object

Why it’s bad: it breaks realism. A kitchen can’t be “lighting-only” if the AI swaps appliances.

 Failure 2 — Workflow logic broke (Kitchen)

Even when the cooktop stayed under the range hood, the AI sometimes:

• pushed a dining table too close to the cooktop
• removed clear working space, making it look decorative, not functional

Why it’s bad: the scene becomes an interior concept image, not a usable room.

Failure 3 — Proportions drift (Bathroom)

Bathroom results often stayed visually similar but failed on realism:

• towels looked oddly placed or distracting
• bathtub height/proportions became unrealistic (not like a normal product)

Why it’s bad: the result feels “AI-ish,” even if the tiles and vanity look correct.

Here are two near-miss examples. They look close at a glance, but small semantic drifts—appliance identity, workflow spacing, and toilet placement—quickly break realism.

Kitchen near-miss (lighting-only, layout-locked): The geometry is close, but the far-left built-in fridge from my reference doesn’t read correctly (it collapses into an under-counter appliance), and the dining table crowds the cooktop under the hood—workflow realism breaks.

Bathroom near-miss example:

Bathroom near-miss (toilet-position lock): Tiles, vanity, mirror, and the shower enclosure stay close, but the toilet drifts into the center beside the vanity instead of staying far right outside the shower—classic “toilet magnet” behavior in bathroom edits.

3 Fixes That Worked Best (My Debugging Playbook)

Fix 1 — Switch to “photo edit” language (Strongest lock)

When “lighting-only” started drifting into redesign, I reframed prompts as:

• “photo edit of the exact same room from the reference image”
• explicitly limiting edits to exposure/white balance/shadows

This reduced “creative remodeling” behavior dramatically.

Kitchen (strong re-generation prompt)

text
minimal modern bright studio kitchen, photo edit of the exact same kitchen from the reference image, lighting-only improvement (exposure/white balance/shadows only),
keep all appliances and their exact positions,
built-in refrigerator on the far left with visible vertical handles (must stay a refrigerator, not a dishwasher),
cooktop directly under the range hood,
no dining table or chairs near the cooktop, keep clear working space in front of the cooktop,
keep lower sink cabinet exactly the same,
no remodeling, no construction, no layout change, exact same camera angle and framing, no new windows, no new doors, no changing tile, no changing flooring
--style raw --ar 4:3 --v 6.1 --iw 2.3 --stylize 10 --chaos 0 --seed 22021

 Fix 2 — Don’t re-roll everything; edit the “near-miss”

If one image is almost correct, re-generating the full scene often introduces new drift.

The highest-success approach is:

• keep the near-miss
• use Editor / Vary (Region) to fix only the broken area

This is especially true for:

• appliance identity issues
•  small workflow corrections
• towel placement problems

Fix 3 — Parameter rule: stylize down, image weight up

When structure started changing, this rule helped:

• lower --stylize to 10–20
•  raise  --iw by +0.1 to +0.3
•  keep --chaos 0 for stability

My short rule:

text

If structure drifts: lower --stylize to 10–20 and raise --iw by +0.1 to +0.3.
Prefer “photo edit of the exact same room” over “redesign”.

Bathroom: Extra Note on “Toilet Magnet” Behavior

Bathrooms have a weird bias: models tend to “want” a toilet in-frame.

If removal prompts keep failing, a practical workaround is:

• keep tiles/grout lines as anchors
•  reframe/pan the scene so the toilet exits the composition
• then patch the empty floor area with “tiles only” edits

It’s not elegant, but it’s often more reliable than fighting the model’s default assumptions.

If you’re trying layout-locked realism, this is the workflow I’d start with.

Before running the next one-variable test, I wanted to document the prompt system and debugging rules I’m using—so the next results are more controlled and realistic.

Previous episode: Clean, Bright, Minimal: A Modern White Kitchen & Bathroom Upgrade (Layout Locked, AI Test)
Next episode: Studio Calm Lighting: Why Bright Rooms Stress You Out + The 3-Layer Renter Fix