We asked three image-to-image models to replace the white-studio background of a uchiwa product photo with a Japanese tatami room. Each model was evaluated on a single prompt and single run. FLUX.2 [dev] and NanoBanana 2 tied at a score of 3.80, with FLUX.2 [dev] being the cheapest. All three outputs were judged to need a reroll because the fan appeared to float over the tatami.
![FLUX.2 [dev] background replacement result](https://pub-568d40d307bf46efb39d070a9ca6a9ee.r2.dev/runs/cmrqoynu8000dx778g0shh249/thumb.webp)
Results
| Model | Score | Verdict | Commercial usability | Cost / image |
|---|---|---|---|---|
| FLUX.2 [dev] | 3.80 | Needs reroll | 2/5 | $0.0120 |
| NanoBanana 2 | 3.80 | Needs reroll | 3/5 | $0.0800 |
| GPT Image 2 | 3.75 | Needs reroll | 3/5 | $0.0800 |
Key takeaways
- FLUX.2 [dev] and NanoBanana 2 tied on score, but FLUX.2 [dev] is the cheapest per image.
- All three outputs had grounding/shadow inconsistency: the uchiwa looked like it was floating on the tatami.
- This is a single-prompt, single-run exploratory comparison; results for other products or backgrounds may differ.
- For production use, expect to reroll or manually composite the output.
Verdict
None of the three models produced a fully usable image on the first try. FLUX.2 [dev] is the cheapest option to iterate with, but all results need a reroll or post-processing.
Note
This is an English summary of the full Japanese benchmark article. Methodology, prompts, all outputs and cost details are documented in the original post.
Read the full article in Japanese →