I recently built Img2Gen, a focused web studio for GPT Image 2 generation.
The idea was simple: I wanted something more practical than a single prompt box. For real creative work, I often need to try multiple aspect ratios, upload reference images, generate several variants, download the results, and keep a history of what worked.
So Img2Gen supports:
- Text-to-image generation
- Reference-image edits with up to 5 uploaded images
- 1, 2, or 4 outputs per request
- Multiple aspect ratios
- Low, medium, and high quality modes
- Daily free generations
- Credit-based paid usage
- Generation history and downloads
- Automatic refunds when a generation fails
The Stack
The app is built with:
- TanStack Start
- React 19
- Cloudflare Workers
- Cloudflare D1 with Drizzle ORM
- Cloudflare R2 for generated image storage
- Better Auth
- Stripe / Creem for payments
- A server-side GPT Image 2 generation provider
I wanted the whole thing to run close to the edge, avoid a traditional server, and keep the architecture small enough that I could iterate quickly.
Why Cloudflare Workers?
Image generation itself is slow compared with a normal web request, so the main app has to handle queued jobs, polling, persistence, and failures cleanly.
The flow looks roughly like this:
- User submits a prompt and optional reference images
- The server validates the request and checks moderation
- Free daily usage or credits are reserved
- A generation job is submitted to the provider
- The client polls for status
- Successful images are downloaded, moderated, and persisted to R2
- The generation record is saved in D1
- Credits are refunded if the job fails
The hardest part was not the UI. It was making the generation workflow feel reliable when the underlying job can take a while or fail halfway through.
Credit Handling
I ended up using a credit model because image quality and output count change the cost.
Currently:
- Low quality: 1 credit per image
- Medium quality: 4 credits per image
- High quality: 16 credits per image
Free users get daily generations, and paid users can use monthly credits or credit packs.
One small detail I cared about: failed generations should not silently burn credits. If a job fails after credits are reserved, the system refunds them automatically.
Reference Images
Reference-image editing was one of the main reasons I built this.
Img2Gen accepts PNG, JPG, and WebP references, up to 5 images per request. This makes it useful for product shots, ad creatives, portraits, style exploration, and iterative edits where you want the output to stay anchored to existing visuals.
What I Learned
A few takeaways from building it:
- AI image apps need good job recovery, not just good prompts.
- Credit systems need to be boring and predictable.
- Storing generated outputs matters because provider URLs are not always a long-term product experience.
- A focused workflow can be more useful than exposing every possible model setting.
- Edge/serverless architectures work well here, but long-running generation flows need careful state design.
Try It
You can try it here:
I would love feedback from other builders, especially on:
- Whether the generation flow feels clear
- Whether the credit model is understandable
- What reference-image workflows you would expect from a tool like this
- Any architecture suggestions for long-running AI generation jobs on serverless platforms