ChatGPT Images 2.5 Flare: Gets the Diagram Right, Gets the Body Wrong

• —ChatGPT Images 2.5 Flare lands 2nd overall on our Human Creativity Benchmark leaderboard, almost on par with Muse Image and ahead of ChatGPT Images 2.0 and Nano Banana Pro.
• —In our testing, Flare excels at material and lighting effects and at isolated product scenes. It leads two of the three campaigns at every design stage: ideation, mockup, and refinement. One designer said "The output on the right (Flare) feels more finished and dimensional. The bottle reads more clearly, the glass caustics are stronger, and the overall lighting better matches the approved campaign direction."
• —Flare struggles with humans, wearables, unusual perspectives, and prompts with many spatial constraints. One campaign, Cairn, asks for "a hiker mid-stride over the camera ... boot tread and the pack's base looming in the near plane." The prompt produced an output with the person's head turned backwards. Designers said "guy is breaking his neck," "body position looks twisted," and that the head and pose looked unnatural.
• —Flare leads two of the four evaluation dimensions: usability and prompt adherence.
• —It trails Muse Image slightly on general preference and visual appeal, which comes down to the one campaign where Flare trailed.
• —We also noticed Flare produces more anatomically accurate diagrams. It uses glyphs to explain instructions, and it conveys information with less text and clearer categorization.

Head-to-head win rate against each opponent in Refinement.