Artificial Analysis Exposes Image Drift Over 30 Edits

Artificial Analysis evaluated four major AI image tools—Ideogram 4.5, OpenAI GPT Image 2.5 Sunburst, Black Forest Labs FLUX 3, and Google Nano Banana 2.1—by applying 30 sequential revisions to a single real estate photograph. The test applied consecutive prompts including adding furniture, lighting a fireplace, painting walls, and shifting daylight to twilight. While standard single-edit leaderboards place GPT Image 2.5 Sunburst Max at the top with an Elo score of 1,197 (ahead of Grok Imagine Image 2.0 at 1,155 and Microsoft MAI-Image-2.6 at 1,150), the multi-turn experiment exposed severe pixel drift. Sunburst repeatedly re-rendered large portions of the frame, leaving only about 20% of pixels intact during small localized adjustments.

In contrast, local copy-and-patch architectures demonstrated much higher scene preservation over extended editing sessions. Both Ideogram 4.5 and FLUX 3 retained at least 95% of unchanged pixels during minor edits like inserting flowers. Meanwhile, Google Nano Banana 2.1 restricted edits to local regions but exhibited a gradual background darkening across the 30-turn sequence. These cumulative stability metrics contrast with traditional single-turn rankings, where Ideogram 4.5 debuted lower at #23 for image editing and #34 for text-to-image generation.

For practical implementation, engineers must match model architectures to the expected length of a user session rather than relying solely on single-pass quality scores. Full re-rendering models like Sunburst—which supports up to 4K resolution, 16 reference images, and xhigh or max quality tiers—remain ideal for one-shot relighting and scene-wide overhaul. However, iterative product editing and branded layout revisions require stability-focused models like Ideogram 4.5, which features native 2K resolution, crop-and-stitch editing, up to 4 reference images, and pricing from $0.008 to $0.22 per image across four quality modes.