Uchiwa Cross-Modal Benchmark
A sample of the report delivered through the Private Japan QA service. It crosses Japanese production scenes — uchiwa image generation, I2V video conversion and Japanese text retention — and visualizes reproducible failures.
Summary
6
5
90
text_rendering
Each model was run multiple times under the same prompt and conditions; human reviewers rated outputs on a 5-point scale.
| Scenario | GPT Image 2 | NanoBanana 2 | FLUX.2 | Veo 3.1 | Kling V3 |
|---|---|---|---|---|---|
| Uchiwa only (front) | 4.2 | 3.8 | 4.0 | n/a | n/a |
| Person holding uchiwa | 3.9 | 3.5 | 3.8 | n/a | n/a |
| Uchiwa with Japanese shop name | 2.8 | 3.0 | 2.5 | n/a | n/a |
| Summer festival + uchiwa | 3.6 | 3.2 | 3.1 | n/a | n/a |
| I2V: fanning uchiwa | n/a | n/a | n/a | 3.4 | 3.8 |
| I2V: keep hand and uchiwa | n/a | n/a | n/a | 2.9 | 3.2 |
Disclaimer
This is a sample report. Actual QA engagements define target models, scenarios and evaluation axes before testing and are delivered privately by default.