AI Pin Maker — multi-model AI studio for image, video, and custom pin design (just shipped)

#ai #webdev #showdev #design

Quick intro

I just shipped AI Pin Maker — an all-in-one AI studio for AI image, video, custom enamel pin, and album generation. One workspace, multi-model routing, free starter credits, no signup required to try.

Live at aipinmaker.com in six languages (en / zh / ja / ko / pt-BR / ar).

What it does

Four product surfaces in a single account, sharing the same credit pool:

Surface	What you get
Image	Text-to-image, image-to-image, image edit
Video	Text-to-video, image-to-video, video edit
Pin	Custom enamel pin / badge design, templates, mockup
Album	Baby / pet / couple / growth themed photo album packs

The differentiator isn't any single feature — it's that one prompt session can flow across all four without changing tools or accounts.

Why multi-model routing

A single text-to-image model is a bottleneck. The best photoreal model is often not the best illustrated style model, which is not the best video model, which is not the best motion model. We route per task across 28+ underlying models so each generation hits the model most suited to it.

Current routing covers:

Image generation — Seedream 4.5, Gemini 3 Pro Image
Stylized illustration — Wan 2.7, Flux family
Image-to-video — Seedance 2.0, Kling, Veo 3
Text-to-video — same video stack
Edits / inpainting — image-to-image with reference

The router picks the model from the task type, not from the user — fewer decisions, better defaults.

What I'd actually use it for

If you fit any of these I want your feedback:

Indie maker / small Etsy store designing custom enamel pins, badges, stickers, or merchandise visuals
Social marketer drafting campaign visuals, posters, short promo videos
Family / pet owner trying baby preview albums, pet portraits, growth keepsakes
AI artist prototyping series across image + video in one go

The free tier ships with starter credits — enough to test the full image + video + pin flow without paying anything. Credits required for any generation are shown before you generate (not after), and uploaded source inputs are auto-cleared after 24 hours.

Tech bits (for the devs reading)

A few things that took non-trivial work and might be interesting to others building similar tools:

Edge model routing — each generation request goes through a router that picks the model from a (task type, content category, user tier) tuple. Latency penalty under 50ms vs hardcoded.
Multilingual SEO architecture — localePrefix: 'as-needed' with proper hreflang, canonical, and Organization / Product schema. SoftwareApplication schema lives on the homepage, Product schema on pricing.
Image / video CDN with x-oss-process thumbnails — original assets stay in OSS; everything in UI is on-the-fly resized via query param. Cuts bandwidth without pre-generating sizes.
Standalone Docker output for Next.js 16 + custom runtime tracing.

I'll write up the routing architecture properly in a separate post if there's interest — leave a comment if you want it.