10 rounds of edits on one photo: MAI-Image drifts least (SSIM 0.774), GPT Image redraws everything
booked78 · reddit · 2026-10-20
A Reddit user quantified multi-round edit drift: one living-room photo, 5 image models, 10 identical sequential edits (each round starting from the previous output), with the never-mentioned bookshelf as a control area measured by SSIM against the original. All 50 edits ran through one API via the Atlas Cloud MCP server; total cost $2.8.
|Model|SSIM after edit 1|After edit 10|Cost/edit|Time|
|---|---|---|---|---|
|MAI-Image 2.6|0.928|0.774 (most stable)|$0.043|40s|
|Nano Banana 2|0.966 (best single edit)|0.641|$0.08|40s|
|Seedream 5.0 Pro|0.800|0.530|$0.036|1m50s|
|GPT Image 2.5 Sunburst|0.590|0.288|$0.017|40s|
|Qwen Image 3.0 Pro|0.198|0.218|$0.078|1m50s|
- All models applied essentially all 10 edits; the difference is purely collateral damage
- Seedream darkened each round until the room looked like dusk; GPT Image 2.5 redraws the whole frame, reinventing book spines every round; Qwen reframed on the very first edit and stayed low, growing hazier
- Practical takeaway: for 3+ sequential edits, use the slowest-drifting model or stack all changes into one prompt from the original
More from Multimodal
- Manga2Animate: Agentic Script and Storyboard Planning for Faithful Manga Animation — NYCU · 2026-10-20
- TCoVP: Spatiotemporally Grounded Video Chain-of-Thought Improves Reasoning Segmentation — Junwei Wen · 2026-10-20
- ByteDance's DeOverlay: Source-Anchored Flow Matching for Video Overlay Removal with New Benchmark — ByteDance · 2026-10-20
- InfiniDream: VLM-Driven Procedural Generation of Open-Ended Photorealistic 3D Worlds — AlayaLab · 2026-10-20
- Tsinghua's Stateful Harness: Executing Rules, Rendering Pixels for Consistent Interactive World Models — Tsinghua-IVG · 2026-10-20
- 30 seconds of surreal AI video transformations, full prompt shared — umesh_ai · 2026-10-20