GPT-Image-2.6 Hands-on: Native GIF Generation Is Fun, but Face Consistency Regresses and Costs Double
卡尔的AI沃茨 · wechat · 2026-11-07
GPT-Image-2.6 quietly launched, and the author tested it extensively. Core updates: native 2-16 frame seamless looping GIFs, cross-conversation character memory (tag a character and reuse it), local inpainting with tamper resistance, motion grid alignment, leader-line spatial alignment, plus a Flare API mode generating images in as fast as 6 seconds.
10 GIF use cases with full prompts: claymation-style stop motion, meme batches, transparent-channel editing stickers, loading animations, animated PPT data charts, product UI demos, 2.5D exploded-view diagrams, historical timelines, treasure-map animations, and art-deconstruction effects.
Three static upgrades tested: character memory holds a protagonist's face across outfit/scene changes; annotated exploded diagrams have clean, well-placed callouts; it can output pure SVG code usable directly in Figma or the browser, though complex curved illustrations drift.
Verdict: photorealistic quality barely improved over 2.5, face consistency regresses after a few edit rounds, and costs doubled ($0.04 → $0.08 per image; a GIF costs $0.2 per draw). Flare's speed gain is negligible with worse quality. Worth a casual try, not worth rebuilding workflows around.
More from Multimodal
- Local pipeline showcase: MiniMax H3 refMod generations from reference image to 4K upscale — edisson75 · 2026-11-13
- Gaussian splats may be the best way to capture how 3D displays like LKG Portrait look IRL — nptacek · 2026-11-13
- AI-generated 'not canon' clip of Schwarzenegger driving with Michael Jordan — alanskimp · 2026-11-13
- H3-Image API gotcha: server keeps billing on client timeout — set max_retries=0 — xiaohu · 2026-11-13
- MiniMax open-sources H3-Image: native 16MP output, up to 16 reference images — xiaohu · 2026-11-13
- Daily AI painting made with GPT-6 Images 2.6 shows painterly winter village — DeryaTR_ · 2026-11-13