/assets/_shared/backgrounds/landing-page.v1.jpg)
Every week another agent posts a Midjourney or DALL·E render and asks why their MLS got flagged. The answer is simple — those models were trained to make beautiful images, not to keep your building the same building. Edensign was trained for the opposite job.
/assets/_shared/icons/upload.v5.png){kind=icon}
Side by side
Four common edits an agent might want. Generic image AI on the left, architecture-trained AI on the right.
Stage an empty living room
Generic image AI
Repaints the walls beige, replaces the floor with hardwood, changes the ceiling height, swaps the windows for French doors. Beautiful — and not your house.
Edensign
Drops furniture into the existing room. Walls, floors, windows, ceiling, baseboards untouched. Only the staged content is new.
Change a daytime exterior to dusk
Generic image AI
Re-renders the entire facade with a new roofline, generic landscaping, and a stock fence. The mailbox disappears, the garage door changes shape.
Edensign
Relights the existing scene. The same building, the same trees, the same fence — just at dusk, with windows lit and sky transitioned.
Declutter a kitchen counter
Generic image AI
Hallucinates a different kitchen entirely. New cabinets, new backsplash, new island shape. Sometimes adds a window where there was a wall.
Edensign
Removes the cluttered items. The cabinets, counters, backsplash, appliances, and view stay byte-identical.
Preview a wall color change
Generic image AI
Paints the wall — and also moves the wall, re-tiles the floor, and rearranges the lights because it doesn't know which pixels belong to which surface.
Edensign
Applies the new color to the targeted wall plane only. Trim, baseboards, ceiling, floor, furniture, and lighting are preserved.
Failure modes
Real outputs from real models on real listing prompts. We track these because we test against them every release.
Asked to "stage this empty living room," Midjourney often regenerates the room — different wall color, different floor, different windows, different ceiling. The composition is great. The building is not yours.
Why it matters
MLS associations require listing photos to depict the actual property. A regenerated room is a misrepresentation, not an edit.
Trained on the internet, DALL·E learned that "a kitchen has a window over the sink" and "a master bedroom has a French door to the garden." It will silently add them — even if your kitchen has neither.
Why it matters
A buyer who tours the home finds a wall where the listing showed a window. Trust is gone. The complaint goes to your broker.
Without an architecture-aware loss, edges bend. Wall-to-ceiling lines curve, baseboards drift, tile grids warp. The image is photorealistic-ish but geometrically incoherent.
Why it matters
Photographers and stagers spot it immediately. Buyers feel "something is off" even if they cannot name it.
Asked to "describe and improve this listing photo," ChatGPT will invent a fireplace, a built-in bookshelf, a view of the ocean — none of which exist in your image. The render then matches the invented description.
Why it matters
You ship a beautiful render of a property that does not exist. Same legal exposure as a fabricated listing description.
No. The failure is in the model, not the prompt. Midjourney has no concept of "the same building." Every render is a fresh creation that approximates the prompt — there is no source-image constraint that says "wall pixels here must equal wall pixels there." Better prompts give you more beautiful inventions, not faithful edits.
Closer — they constrain the edit to a masked region — but the underlying model is still a general-purpose image model. It will happily change a wall texture or window shape inside the mask because it has no real-estate-specific loss term. For one-off cosmetic touch-ups on personal photos, Firefly is great. For listing-faithful editing at MLS scale, it's the wrong tool.
No. We train and run our own model. Every output is generated by a model that we trained on a proprietary corpus of architectural and listing photography, with explicit losses that penalize structural drift. We do not call OpenAI, Stability, or Midjourney APIs for image generation.
They keep getting better at images. But "better images" and "faithful listing edits" are different problems. As long as the training objective rewards beauty over identity, the gap on structural fidelity remains. We retrain monthly to stay ahead on the metrics that matter for real estate.
Send them the architectural-ai page — it includes our methodology, benchmark numbers, and dataset provenance. We can also deliver a model card under NDA for brokerage and enterprise customers.
Read next
Pillar
Dataset, training stages, benchmarks, and the architectural reasons our model preserves your building.
ReadEngineering guarantee
How element-preservation constraints keep walls, floors, windows, and built-ins identical between source and output.
ReadMLS policy
The compliance side — what MLS associations require, how Edensign meets the bar by default.
Read