Meta released Muse Image this week so I ran it against OpenAI's gpt-image-2 and Google's Nano Banana 2. I used the same source duck image and the same edit instructions prompt for every model (unchanged β†’ blue β†’ face away β†’ glass β†’ wireframe β†’ hat-on-ball β†’ "FRENZY" text β†’ standing on a mirror with a correct reflection). The transformations go from easy on the left and gradually get harder. I ran 3 runs per model. Each model was then scored using a fixed 27-point rubric. One of thes...