Fine Photo Gen

Put your best foot forward and always look fine with Fine Photo Gen. Why pay a photographer hundreds of dollars for headshots when AI does it for $40? This app makes you look good in every scenario.

Comment

The Inspiration

It all started with a tweet. I was scrolling through my feed when I saw a post from a solo developer claiming they were making over a million dollars a year with a personalized AI-photo service. I stopped scrolling immediately. The industry heavyweights like Midjourney were focused on general-purpose image generation, but the niche of creating custom, fine-tuned models for individuals was wide open. I saw a clear gap in the market and a proven business model. My first thought was, “Why not me?” If he could do it, I could do it too. I've been waiting for the right AI idea.

How I Built It: The Tech Stack

I knew I needed a modern, robust, and scalable stack to pull this off. I chose technologies that would allow for rapid development without sacrificing quality or performance.

Layer Tools & Libraries Notes
Frontend Next.js 14 (App Router), Tailwind CSS, Shadcn/ui Rapid UI development
  Zustand Lightweight state management
  React Dropzone Seamless file uploads
Backend Node.js & TypeScript via Next.js API routes Unified front- and back-end
Database PostgreSQL + Prisma ORM Type-safe queries & migrations
AI & Cloud Services Replicate – LoRA training & inference pipeline
Cloudflare R2 – Store zipped training sets & generated images
HuggingFace – Host model files post-training		 Core of the product
Payments Stripe Stripe for subscriptions
Authentication NextAuth.js Secure, full-featured auth
Testing Jest + React Testing Library Unit & integration tests

Challenges I Faced

My biggest hurdle was a misleading blog post. The original plan:

  1. Train LoRA models on Replicate.
  2. Upload fine-tuned weights to a HuggingFace repo.
  3. Use Together.ai for fast, cheap inference.

After a day or two of work I discovered Together.ai’s support for custom HuggingFace LoRAs simply didn’t exist. I had to scrap the flow and re-architect the AI pipeline, ultimately keeping both training and inference on Replicate. This required a major rewrite of my service classes and logic—a tough reminder to verify marketing claims before betting on them.

What I Learned

  • Infrastructure matters. Manually configuring Nginx reverse proxies across services was painful. I’ll invest in a dedicated configuration tool or Infrastructure-as-Code (e.g., Terraform) moving forward. I always wanted to host models directly instead of relying on an external service, but I was forced to because my server doesn't have a dedicated GPU. Now I'm happy to offload whatever I can off of my server as long as I can afford it. That won't stop me from building innovative apps any longer.
  • Orchestrating async workflows is tricky. The training pipeline—validating & zipping images, uploading to R2, calling Replicate, polling status, updating the DB—forced me to design robust state management and error handling so users always know where their job stands.
  • Validate early, refactor fast. Missteps in third-party tooling can cost precious time; quick pivots keep momentum alive.

Built With

Share this project:

Updates