A design director in a skill: Chinese architecture photos become real posters

What it does You hand it a photo of a temple, an old street, a teahouse, or a local dish, and it produces a fully composed editorial poster via GPT Image. It reads the photo’s composition and identity, picks one dominant visual reference from built-in recipes, designs Chinese display type, and hands the model the original image, the reference, and a sparse layout plate so everything gets redrawn in one coherent visual world.

The interesting bit The pipeline is unusually disciplined for an image-generation tool: it formulates creative direction before looking at references (to avoid “more references, more mediocre output”), protects identity anchors like roof slopes, flying eaves, and plaques so your real building doesn’t get optimized into a different building, and runs deterministic pre-flight gates — font cmap coverage, collision checks, vertical text order — before spending an expensive model call. After generation it does exactly one cheap read-back to diagnose problems, then hands control back to you instead of silently burning another generation.

Key highlights

• Recognizes composition types (frontal, low-angle, framed view, detail) and fixes only one real compositional problem per poster
• Multi-image mode extracts subjects from several photos and rebuilds them in one shared scene with common perspective and light — not a five-grid collage
• Display type is designed with stroke contrast, weight distribution, and terminal treatment; small text uses a separate regular family for hierarchy
• Optional extension: a 3×3 video storyboard plus prompts ready for a video model
• Only verified, photo-visible, or user-confirmed facts go on the poster — no invented history

Caveats

• Needs an agent that supports Skills and image generation, plus Python 3.10+ with Pillow, NumPy, OpenCV, and fontTools — it’s not a standalone app
• Explicitly not for e-commerce product shots, routine retouching, or fictional architecture with no real subject

Verdict Worth a look if you produce editorial or cultural-heritage visuals with Chinese typography and have a capable agent handy. If you just want exposure fixes and a sky swap, this is a very elaborate way to not do that.