Curious Refuge ranks GPT Image 2 above Meta Muse in character tests

• GPT Image 2 led the bakeoff on character consistency and realism, while CuriousRefuge's verdict said Muse Image still held up.
• Muse Image's clearest win condition was structured information design: CuriousRefuge's verdict singled out accurate infographics, and the baseball chart prompt tested hit breakdowns and home-run leaders.
• The test set mapped to working creative jobs: the poster prompt asked for theatrical typography, the cinematic still prompt asked for identity preservation, and the edit prompt asked for composition-preserving glasses placement.
• The debate moved beyond raw model quality because petergyang's benchmark question asked how anyone benchmarks design taste.

Meta's Muse Image launch post describes it as the first image generation model from Meta Superintelligence Labs, with multi-photo blending, presets, sketch edits, Instagram and WhatsApp surfaces, and a July 10 update removing public-Instagram @mention references after feedback. OpenAI's GPT Image 2 model page frames its rival as a state of the art model for fast, high-quality generation and editing with high-fidelity image inputs. Curious Refuge's prompts were production-shaped: character sheets, infographics, posters, cinematic stills, edits, and covers.

Meta introduced Muse Image on July 7 as a Meta AI image model built for conversational prompting, photo blending, direct image markup, and sharing into Meta apps. The same launch post says Muse Image is free for everyday creation, available through subscription plans for heavier use, and coming to Advantage+ creative for advertisers.

That is the model in CuriousRefuge's comparison with GPT Image 2, which called Muse brand new and tested it against what the account considered the strongest current image model.

The thread used six creator-facing tasks rather than one pretty prompt.

• Character sheet: CuriousRefuge's first prompt asked for front, side, back, three expressions, gear closeups, consistent design, and a non-cartoon concept-art style.
• Infographic: CuriousRefuge's second prompt asked for a hit-type breakdown and home-run leaders from a chart.
• Poster: CuriousRefuge's third prompt asked for a theatrical sci-fi poster with the title “THE LAST SIGNAL” and a tagline.
• Cinematic still: CuriousRefuge's fourth prompt asked the models to preserve a reference character inside an abandoned 1970s movie theater.
• Image edit: CuriousRefuge's fifth prompt asked for glasses placement, a mid-century modern room, and the same general composition.
• Book cover: CuriousRefuge's sixth prompt asked for a premium nonfiction cover with exact title, subtitle, author, and no random extra text.

The character prompt stacked identity, costume, expressions, accessories, and multiple body views into one sheet. In CuriousRefuge's verdict, GPT Image 2 won on character consistency and overall realism.

Muse Image's strongest showing was structured graphics. CuriousRefuge's verdict said Muse produced accurate infographics, even if GPT Image 2's visuals looked more polished.

Meta's launch copy makes the same bet: the Muse Image post says the model can render clean text, build how-to guides, produce detailed infographics, and even generate a functional QR code.

The middle of the test moved from layout into film language and compositing.

• The poster prompt demanded premium sci-fi drama mood, title typography, tagline control, and no extra random text.
• The cinematic-still prompt demanded the same face, wardrobe, and mood from a reference image, then placed the character in a 16:9 theater scene.
• The edit prompt in CuriousRefuge's image-edit test checked whether the model could add glasses and relocate the subject without breaking composition or style.

The cover prompt was a text-rendering trap: title, subtitle, author name, premium negative space, and a ban on misspelled words or random text. OpenAI's image prompting guide calls out reliable text rendering with crisp lettering and consistent layout, which is why this prompt belonged in the bakeoff.

petergyang's reply landed on the hard part: how do you benchmark design taste? The prompts asked for “premium,” “theatrical,” “A24,” “not cartoonish,” and “airport bookstore,” but those scores still came from human visual judgment, not a standard metric.