The usual mental model for locking a reference image is that you're preserving what you specified. You described a character, generated her, picked the best result, and now you're keeping that description steady across later images. A recent test suggests the reference preserves something broader than the description: everything that happened to be in the picture, whether anyone asked for it or not.
What happened
The character was defined with a deliberately small description: hair, eyes, a two-color outfit, one signature accessory. Three variants were generated from it. One of them, the base version, came back with a small extra hair clip on the left side of the bangs. Nothing in the prompt mentioned it. It didn't show up in the other two variants.
That base version was picked as the master, for reasons that had nothing to do with the clip: the contrast was good and the line work was clean. The master then became the reference for everything downstream. When it was later fed to a second tool that generates a stream-style layout, with the master image as its only input, the clip was there in the result. The prompt for that step never mentioned a clip. The reference did.
Why it works that way
A reference image doesn't contain a list of intended traits and a separate list of accidents. It's just the picture. A model reading identity off that picture treats anything visually distinctive as a candidate for preservation, and it has no way to tell "this was in the brief" from "this just happened."
So the written description and the master image can quietly diverge. The description says one clip's worth of accessories. The image says one more than that. Once the image is the reference, the image wins.
What to do about it
Inspect the master for incidental details, not just the intended ones. Picking a master by checking the specified features is the obvious move. It's also worth scanning for anything extra that you'd be stuck with.
Fix unwanted accidents before locking, not after. At the master stage, removing a stray detail costs one regeneration. After the image has been used as the input to everything else, the detail has been inherited by every downstream output.
Write down the good accidents. If a detail works, it's now part of the design, and nothing in your original written description records it. A handoff note to an artist or rigger that lists only the specified traits would leave it out, even though the actual character has it.
The limits of this
One test. It doesn't show that an incidental detail always survives; the tool's own documentation says small details like accessory placement can drift, so persistence here was an observation, not a guarantee. It also doesn't show how the detail would hold up under more aggressive changes than a plain background swap or a slight angle shift.
The general takeaway
When the image is the source of truth, the description is a summary of it, and summaries leave things out. If something matters enough to protect, it needs to be in the written handoff. If something shouldn't be there, the time to catch it is before the image becomes the reference.
Top comments (0)