--- title: "Face and Character Consistency Test (Preregistered) | Made Research" description: "Nano Banana vs Nano Banana 2 vs Nano Banana Pro in a preregistered, blinded benchmark: 3 runs per prompt, published failure rates, no provider sponsorship." url: https://www.made.app/research/face-character-consistency-2026-q3 --- [Research](https://www.made.app/research) / Face and Character Consistency Test # Face and Character Consistency Test Preregistered The protocol below is frozen. No outputs have been generated or scored yet, so this page makes no quality claims — it commits to how the claims will be made. Protocol frozen 2026-08-13 · methodology v1.0 · 3 runs per prompt per model ## Models under test Nano Banana 4 credits / image [Full pricing](https://www.made.app/price-sheet/nano-banana) Nano Banana 2 from 6 credits / image [Full pricing](https://www.made.app/price-sheet/nano-banana-2) Nano Banana Pro from 15 credits / image [Full pricing](https://www.made.app/price-sheet/nano-banana-pro) ## Metrics | Metric | How it's scored | Scale | Better | | --- | --- | --- | --- | | Blinded face similarity | human rating | 1-5 | Higher | | Overall character consistency | human rating | 1-5 | Higher | | Scene instruction adherence | human rating | 1-5 | Higher | | Anatomical artifact count | measured and rated | count | Lower | ## Test protocol **Aspect ratio** 4:5 **Reference policy** Use the same synthetic adult headshot for every model, prompt, and run. **Output count** 1 ## Rating & integrity - 2 blinded raters score every output. Randomize model labels and output order independently for every prompt. - Publish both scores, their absolute difference, and the adjudication rate. - Count provider errors, safety refusals, corrupt media, and empty outputs as failures; do not silently rerun beyond the preregistered run count. - Exclude only confirmed infrastructure failures affecting every compared model; record the reason and original run ID. - Cherry-picking prohibited; no sponsored compute; no provider review before publication. ## Results — not yet published This study publishes no scores until every preregistered run is complete, both blinded raters have scored every output, and two reviewers have signed off. When that happens the summary tables and the full raw run data will appear here — including failure rates and rater disagreement, not just the winning numbers. ## Frozen prompt set: character-consistency-2026-q3-v1 All 20 image-generation-with-reference prompts, frozen 2026-08-13 at version 2026-q3-v1. Every model runs every prompt 3 times. Download as [JSON](https://www.made.app/research-data/character-consistency-2026-q3-v1.json). 1. Place the synthetic adult reference subject in a bright ceramic studio wearing a charcoal apron, medium portrait, neutral expression, realistic window light. 2. Show the same synthetic adult subject reading at a cafe table in a navy sweater, three-quarter portrait, soft morning light. 3. Show the same synthetic adult subject walking on a quiet city sidewalk in a tan coat, full-body editorial photograph, overcast daylight. 4. Show the same synthetic adult subject presenting beside a blank whiteboard in a modern office, waist-up, even commercial lighting. 5. Show the same synthetic adult subject cooking at a home kitchen counter in a green shirt, candid medium shot, warm evening light. 6. Show the same synthetic adult subject seated on a red train seat, looking through the window, natural travel photograph. 7. Show the same synthetic adult subject holding an unbranded blue product box in a gray studio, centered commercial portrait. 8. Show the same synthetic adult subject outdoors at blue hour wearing a black rain jacket, close portrait, wet street bokeh. 9. Show the same synthetic adult subject sketching in a museum gallery, side profile, quiet editorial composition. 10. Show the same synthetic adult subject standing on a rooftop at golden hour in an ivory jacket, waist-up portrait, city softly blurred. 11. Show the same synthetic adult subject arranging flowers at a market stall, documentary medium shot, natural color. 12. Show the same synthetic adult subject sitting on a library floor beside stacked books, relaxed full-body portrait, soft daylight. 13. Show the same synthetic adult subject cycling slowly on a coastal path with a matte helmet, wide lifestyle photograph. 14. Show the same synthetic adult subject holding a ceramic cup near a sunlit window, close portrait, realistic skin texture. 15. Show the same synthetic adult subject inspecting a camera in a minimal studio, three-quarter portrait, controlled softbox light. 16. Show the same synthetic adult subject wearing a red scarf in falling snow, centered close portrait, cold natural light. 17. Show the same synthetic adult subject watering plants in a glass greenhouse, full-body editorial photograph, diffused light. 18. Show the same synthetic adult subject seated at a piano in a dark rehearsal room, side-lit medium portrait. 19. Show the same synthetic adult subject carrying a canvas tote through an outdoor market, candid full-body photograph. 20. Show the same synthetic adult subject facing camera against a plain warm-gray background, neutral wardrobe, reference-style headshot. ## Limitations - One synthetic identity does not represent all faces, skin tones, ages, hairstyles, or reference qualities. - Similarity ratings can reward an attractive but incorrect reconstruction, so scene adherence and artifact scores remain separate. ## Disclosures - The benchmark uses a Made-owned synthetic adult identity and no private user photos. - No provider sponsors this study. All studies in the program are listed on the [research index](https://www.made.app/research); the models under test are priced on the [price sheet](https://www.made.app/price-sheet). --- Markdown version of https://www.made.app/research/face-character-consistency-2026-q3. Every public Made page has one at its URL plus .md (https://www.made.app/research/face-character-consistency-2026-q3.md). Site summary for AI assistants: https://www.made.app/llms.txt