- Reference-guided editing
- Localized revisions
- Multi-reference workflows
Tested in mirror-selfie and moderation comparisons
Nano Banana 2
A strong reference-combination and editing route with useful evidence in AI Photo's mirror-selfie benchmark. Its task-level completion and moderation behavior remain visible row by row.
- Safety blocks vary by prompt
- Mirror logic remains difficult
- Not represented in the current Arena collection
Inspectable evidence
Tests containing Nano Banana 2
ComparisonsBest AI Models for Realistic Mirror Selfies: 15 Same-Input Comparisons
Fifteen visible comparisons from 18 tested mirror-selfie prompts: 39 retained outputs and 15 recorded blocks across Nano Banana, Nano Banana 2 Lite, and Nano Banana 2.
Open test →ComparisonsWhich AI Image Generator Blocks More? 9-Attempt Same-Input TestNano Banana 2, GPT Image 2 Medium, and Seedream 4.5 tested on the same three adult-fashion prompts: nine attempts, seven outputs, and two visible safety blocks.
Open test →ComparisonsNano Banana 2 vs Seedream vs GPT Image 2 vs Grok: 14-Prompt Portrait ComparisonFourteen adult portrait prompts with side-by-side outputs from Nano Banana 2, Seedream 4.5, GPT Image 2 Medium, Grok, and Seedream V5 Lite.
Open test →Related tasks
Related task guides
Task guideAI Image Editing
Judge edits by what changed, what stayed fixed, and how many attempts were required—not by whether the final image looks attractive in isolation.
Open task guide →Task guideAI Portrait RealismCompare skin texture, anatomy, camera logic, lighting, and the difference between a polished image and a believable photograph.
Open task guide →Evidence guideAI Image ModerationCompare documented outcomes from the same permitted adult-fashion prompts without extrapolating a platform-wide policy from one request.
Open task guide →