50 sequential image edits tested: Seedream drifts only 36.6, MiniMax collapses at edit 8
aziz4ai · x · 2026-11-16
Blogger aziz4ai ran a "50 sequential edits" stress test on four image-editing models: same source image, same prompts, each edit applied to the previous output, with instructions that reverse halfway (e.g. turn the hoodie red, then black again) so the final image should return close to the original. First attempt only, no manual fixes, drift measured in code (lower is better).
- Seedream 5.5: drift 36.6 — the only model that came back near the original, preserving faces and lighting to the last step, and the cheapest.
- Nano Banana 2.1: drift 74.7 — great early on, but gradually added unrequested changes, inventing wall paintings by the end.
- GPT Image 2.6: drift 76.4 — most precise at executing each individual edit, but shifts the camera and adds noise each step; most expensive.
- MiniMax H3: drift 97 — collapsed around the 8th edit with fully blurred faces; it also force-rewrites prompts automatically with no way to disable, hurting results.
Takeaway: for many rounds of edits on one image, Seedream is the most stable; for one or two edits, Nano Banana and GPT are excellent; the more edits stack, the more damage — better to restart from the original than keep building on the latest result. Full steps shown in the video.
More from Multimodal
- AI-generated VHS zombie short stuns with photoreal rainy-night close-up — brythehuman · 2026-11-16
- Reddit user shares MiniMax image workflow for fixing broken hands — Friendly-Fig-6015 · 2026-11-16
- Pro artist tests Minimax Image Edit: sketch-to-realism beats Nano Banana, weak at composition control — GrungeWerX · 2026-11-16
- Player has Gemini judge an image contest — and Moss wins every round — FloppyGhost0815 · 2026-11-16
- Art made with Claude Fable 5.2 draws praise for its style — repligate · 2026-11-16
- AI brings an old PC game intro to life with moody foggy wasteland visuals — seppe0815 · 2026-11-16