This is a limitation of today's image generation technology and applies to every AI image tool on the market.
Image models don't "edit" a photo the way Photoshop does. They regenerate it. Every pixel is redrawn from scratch based on the photo plus the instructions. That means the model is always making judgment calls about what it thinks it's seeing — and in a redraw, a doorway can drift, a window can resize, a countertop can shift tone. Instructions steer that process strongly, but they don't lock it. It's guidance, not a constraint. The same request can come out perfect nine times and wrong the tenth.
We already have guardrails
Every photo edit and virtual staging request automatically carries built-in rules telling the image model to preserve architecture, room dimensions, window and door placement, and to only remove clutter or add furnishings. Those rules are baked into the product — they're on every single edit whether or not you've written a Training Moment. And we're continually tuning them: as the underlying models improve and as we learn from real edits, we tighten those instructions and roll the improvements out to everyone automatically.
Why a Training Moment can't fix how an image comes out
Training Moments coach the assistant's behavior and words — what it agrees to do, how it responds, what it discloses. They're read by the assistant, not by the image model that draws the picture. So a Training Moment saying "don't move the doorways" does shape how BrokerBot talks to you about an edit, but it isn't what governs the pixels. Preservation instructions like that already go directly to the image model on every edit — restating them in a Training Moment doesn't strengthen the instruction the model receives, because it's already receiving it.
What to do instead
Start a new chat. Long threads pile up context — old photos, earlier edits, prior instructions — and the model starts blending them together. A fresh chat for each image gives it a clean slate and noticeably better results.
If it drifts, just ask again — regenerating may fix it.
If you keep hitting the same issue, let us know through the Support tab so we can investigate and improve it.
Always disclose virtual staging.
Looking ahead
This technology is improving fast, and we ship improvements to it continuously — what's inconsistent today gets more reliable over time. In the meantime, don't forget your other integrations: Canva's Magic Studio tools are great for precise touch-ups, and pairing BrokerBot's generation with Canva's editing gives you the best of both.
Thanks for building with us,
The BrokerBot Team