OPEN LINK ↗
// 1h agoBENCHMARK RESULT
Artificial Analysis Exposes Image-Edit Drift
Artificial Analysis tested four frontier image-editing models across 30 consecutive edits of the same photo, measuring how much of the original scene survived outside the requested changes. Ideogram 4.5 and FLUX 3 preserved at least 95% of pixels during small edits, while GPT Image 2.5 Sunburst and Nano Banana 2.1 introduced substantially more cumulative drift.
// ANALYSIS
This benchmark measures the failure mode that matters most in real creative workflows: whether approved details survive iteration, not just whether one edit looks impressive.
- –Local-patching models currently look better suited to product photography, advertising revisions, and layout-sensitive work.
- –GPT Image 2.5 Sunburst appears optimized for broad reinterpretation, making it stronger for one-shot restyling than long edit chains.
- –Nano Banana 2.1 reportedly follows individual instructions well but gradually shifts brightness and background details over time.
- –Pixel preservation alone is incomplete; benchmarks should score edit compliance and protected-region stability separately.
- –Developers should test representative multi-step sessions instead of relying solely on single-edit leaderboards.
// TAGS
artificial-analysisbenchmarkevaluationimage-genvisionmultimodal
DISCOVERED
1h ago
2026-10-11
PUBLISHED
1h ago
2026-10-11
RELEVANCE
8/ 10
AUTHOR
TraffAlex
