I've spent the past few months working on AI image generation tooling, and a lot of friends have asked me what stack I'd use if I were starting fresh in 2026. Here's the honest answer, with the gotchas I hit along the way.
This isn't theoretical — it's based on what I've seen working in production at Nano AI (a free Nano Banana Pro workbench) and what I'd build into the next product.
The Stack
| Layer | Choice | Why |
|---|---|---|
| Frontend | Next.js 15 (App Router) + Tailwind | Boring, fast, server components for SEO |
| Auth | NextAuth | OAuth (Google/GitHub) covers 95% of users |
| DB | PostgreSQL (Supabase or Neon) | Free tier covers MVP, scales when needed |
| Image Model | Nano Banana Pro / SDXL fallback | Cost per image determines unit economics |
| Storage | Cloudflare R2 | S3-compatible, 10x cheaper for image-heavy apps |
| Payment | Stripe | Don't reinvent this |
| Hosting | Vercel (frontend) + Cloudflare Workers (image proxy) | Edge for cold starts |
| Analytics | PostHog (self-hosted) | Privacy-friendly, full event control |
Total monthly fixed cost at MVP scale: ~$30-50. The real cost is per-image inference.
The Critical Decision: Where Do You Run the Model?
This is the make-or-break choice for an AI image SaaS. Three options:
Option 1: Hosted API (OpenAI, Stability)
- ✅ Zero ops
- ❌ Per-image cost is high — kills free tiers
- ❌ Vendor lock-in
Option 2: Self-host on RunPod / Modal / Replicate
- ✅ Cheaper per image
- ❌ Cold start pain (10-30s for first request after idle)
- ❌ More ops
Option 3: Use a small efficient model via cheap inference provider
- ✅ Best per-image cost (Nano Banana Pro is in the $0.001-0.003 range)
- ✅ Sub-second inference
- ❌ Quality ceiling lower than DALL-E 3 for complex prompts
For a freemium SaaS, Option 3 is the only thing that makes the unit economics work if you offer a free tier.
I wrote about the actual cost breakdown across providers — the gap between the most expensive and cheapest option is 40x.
Free Tier Strategy
This is the part most builders get wrong. Here's what works:
Don't offer "unlimited free"
You'll get bot abuse. Within a week you'll be paying for someone else's bulk generation.
Don't offer "10 free per month"
Visitors will hit the limit on their first session, never come back.
Do offer "10 free per day"
Resets daily. Keeps engagement. Shows the product enough times to convert.
Implementation: Redis or Postgres counter keyed by IP + cookie. Reset every 24h.
The Bot Abuse Problem
Within the first week of putting up an AI image generator, you'll have:
- Scrapers hitting your
/api/generateendpoint - Users with disposable emails creating 100 accounts
- Distributed attacks from residential proxies
Defense in depth:
- Cloudflare Turnstile on signup (free, lighter than reCAPTCHA)
- Rate limit by IP + device fingerprint (not just IP — proxies are cheap)
- Image hash deduplication — if the same prompt+seed already ran today, return the cached image
- Behavioral checks — humans take >0.5s to hit "generate"; bots don't
What I'd Skip in 2026
A few things that seemed essential a year ago but really aren't for an MVP:
- Custom auth flow — NextAuth handles 95% of cases, don't waste time building from scratch
- Microservices — one Next.js monolith handles 10K req/min easily on Vercel
- Proprietary AI model — you don't need to fine-tune your own model in 2026; the open-weight ecosystem is good enough for most use cases
- Real-time collab features — adds complexity, rarely needed for image gen apps
Lessons (the hard-won ones)
Cost dictates everything. Get the per-image cost down before adding features. Otherwise you're building a money-losing machine.
Cache aggressively. A surprising number of "AI generations" are people typing similar prompts. Cache by prompt+seed.
The model's quality ceiling won't matter if your prompt UX is bad. Build a great prompt-input experience (suggestions, examples, history) and a 7/10 model beats a 9/10 model with a bad UX.
Free tier conversion is brutal in this space. Average is ~1-2% (paid). If you're below 0.5%, look at your free-to-paid friction, not your pricing.
Image hosting is its own line item. Generating a 1024x1024 PNG is one cost. Storing/serving it for 1000 visitors is another. Use R2 or B2.
Try the Stack in Action
If you want to see one concrete instance of this stack running in production, Nano AI is exactly the thing — Next.js + Nano Banana Pro + Cloudflare R2 storage.
Try the workbench, look at the load behavior in DevTools, and you'll get a sense of what "Option 3" feels like in practice.
Closing
The interesting thing about 2026 is that you can ship a real AI product as a single founder for under $50/month in fixed costs. The bottleneck isn't infrastructure or models — it's distribution and prompt UX.
If you're building anything in this space, I'm happy to chat. Drop your project in the comments.
Top comments (0)