Urgent.News

What's breaking now, across thousands of outlets.

Editions

AI

How to Stop an Image Model From Ignoring Your Reference Image

A user uploaded a floor plan, typed "remove the room labels", and got back the same plan — labels still on it. Tried again, same result. The interesting part: the model was not ignoring the instruction. Another line in our own prompt was quietly outranking it. I work on FloorDrafter , an AI floor plan generator. A big part of the product takes a reference image — a photo of a hand sketch, a…

A user uploaded a floor plan and attempted to remove the room labels, but the AI model failed to ignore the instruction. The model was not following the user's reference image. FloorDrafter, an AI floor plan generator, uses a reference image as a starting point and applies a requested change. The challenge lies in balancing the model's two masters: the visual reference and the user's textual instructions.

Several lessons were learned in making the model behave correctly. First, separate the "how to draw" from "what to draw" by splitting style presets into two halves: how to render the image and what annotations to include. Second, write the rules to "only change what is named," not as a list of prohibitions. Prohibitions can lock the user out, preventing them from making necessary changes. Third, explicitly authorize the requested change to avoid the model becoming overly conservative.

Fourth, do not ban existing content; instead, use affirmative clauses that preserve what is already present. For instance, adding a prohibition on text can inadvertently remove user-invented text. Fifth, some constraints must be immune to exemption clauses, such as geometric integrity. When a user wants to change dimensions, the model should not be allowed to duplicate or extend elements without explicit permission.

Lastly, every section that can override constraints should have an exemption included. This ensures that capabilities like removing labels function as intended.

Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.

Read the original at dev.to →

More in AI

Benchmarks Don't Build Great Products. Engineers Do.

The AI industry has a benchmark problem. Not because we have too many benchmarks. Because too many companies treat them like trophies instead of tools.

  • Benchmarks should enhance products, not just market them
  • Engineers should prioritize solving real problems over benchmark optimization
  • Transparency in benchmarking fosters reproducibility and improvement

More from Friday 21 August →