50 sequential edits stress test: Seedream drifts least at 36.6, Nano Banana at 74.7
aziz4ai · x · 2026-11-17
A stress test of four image-editing models: same image, same prompts, 50 chained edits where each round edits the previous output, including round-trip changes (hoodie red then back to black) that should return to the original. No rerolls, drift computed by code.
- Seedream 5.5: drift 36.6 — the only model that ended close to the original, kept faces and lighting to the last step, and was the cheapest
- Nano Banana 2.1: drift 74.7 — great early edits but started unrequested changes, eventually hanging picture frames nobody asked for
- GPT Image 2.6 and MiniMax H3 also tested
Takeaway: cumulative drift over multi-turn editing is a widespread weakness; Seedream leads on stability and price.
Related event: Seedream Drifts Least in 50-Round Image Editing Stress Test of Four Models(2 posts)→
More from Models
- Cua and Scale AI benchmark computer-use agents: 97% non-intrusive on Omarchy vs 40% on Ubuntu — TianbaoX · 2026-11-17
- Vibe math suggests new Astra may already be near 50% on Frontier MazeBench — scaling01 · 2026-11-17
- User slams Google model's personality: unempathic, stubborn, pedantic and impatient — Yamapama · 2026-11-17
- Google Fixes 219 Glitch Tokens in Gemma 4 Tokenizer; GGUF Versions May Lag Behind — TuskNaPrezydenta2020 · 2026-11-17
- FutureSim team: Exa Snapshot fixes leaky backsearch, opens new agent evals — maksym_andr · 2026-11-17
- Moonshot AI Pivots Kimi From Coding Benchmarks to Science Research — The Decoder · 2026-11-17