Replacing objects in images: what the model changes
An image model can make a convincing replacement and still put it through another object or change something nearby. It can also turn a leather sofa beige while leaving leather creases in what should be linen. The work for Muse, our node canvas at SplayFlo , spans 405 edited images across drawn and photorealistic scenes. The outputs came from nine named models or variants and Codex's image tool.…
Image models can convincingly change objects within pictures while altering nearby elements. Muse, a node canvas at SplayFlo, processed 405 images across drawn and photorealistic scenes using nine different models and Codex's image tool. The results came from both specific object descriptions and image references. Overall, 12 out of 15 edits were usable, with Meta Muse being the most successful and cheapest paid model at $0.01 per image.
However, some edits resulted in unusable images, such as a FLUX lamp's base fusing into a sofa and floating above the floor. When a laptop was replaced with a desk and chair, it disappeared in all three attempts, but naming the items in the prompt ensured their presence in six out of six edits.
Text prompts and image references both produced varying results. A description-only request yielded plausible but non-specific outcomes, while an image reference led to 28 sofa and arch results resembling the reference. The text-only sofa had straighter arms and a flatter weave, failing to maintain the looped bouclé weave. Some edits outside the marked area, such as changing the arch color and replacing the room's sofa color, were evident in the photorealistic scenes.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
