The hardest part was the Remove Objects tool. It uses a two-canvas mask editor where you paint over what you want removed.
On mobile, I had to build a custom pinch-to-zoom system with a 5-state machine (idle/drawing/pinching/panning/cooldown) to distinguish between "user is painting a mask" and "user is zooming in."
Getting that right on iOS took more iterations than I'd like to admit.
For the AI pipeline, the server crops only the masked region (with context padding), processes it at 512x512 via LaMa, then feather-blends the result back into the full-resolution original.
Non-masked pixels are completely untouched - zero quality loss outside the edit area.
The standard tools genuinely make no network requests. Easiest way to verify: open any standard tool, turn on airplane mode, and it still works.
That's also why the PWA has offline support for those tools. Both img.tara.vision and our PDF toolkit (pdf.tara.vision) share about 90% of their frontend code.
Same auth, billing, PWA infrastructure, UI components. Building the second product took a fraction of the time because of that.
Happy to answer questions about the architecture, privacy approach, or anything else. Feedback on the mobile UX is especially welcome - that's where I've spent most of my time lately.
Standard tools (Compress, Convert, Resize, Crop & Rotate, Strip Metadata) run entirely in your browser using Canvas API. Open your network tab - zero requests. Your files never leave your device.
AI tools (Remove Background, Remove Objects) run on our own server in Germany. Images are processed in memory and deleted immediately - no storage, no logs, no third parties. We use rembg + u2net for background removal and LaMa for inpainting, running in a Python FastAPI sidecar.
The stack: Vue 3 SPA + Laravel 12 API + PostgreSQL + Python service. PWA with offline support for the standard tools.
Why I built this: Every time I needed to remove a background or compress an image, I'd end up on some site that uploads my files to who-knows-where. For personal photos that's annoying. For client work or medical/legal documents, it's a real liability. I wanted tools I'd trust with my own data.
What's free: All 7 tools, 3 operations/day, no account needed. Pro is /month for 60/day.
Honest limitations:
- The AI models are good but not state-of-the-art (tradeoff for self-hosting on modest hardware)
- 7 tools vs. competitors with 50+. I'm focused on doing fewer things well.
- Single founder, single server. Not enterprise-scale yet.
I'd love feedback on the UX, especially the Remove Objects tool (brush-based mask editor with pinch-to-zoom on mobile). Available in EN, DE, FR, ES, IT, MK.
Founder here. Some behind-the-scenes context:
The hardest part was the Remove Objects tool. It uses a two-canvas mask editor where you paint over what you want removed. On mobile, I had to build a custom pinch-to-zoom system with a 5-state machine (idle/drawing/pinching/panning/cooldown) to distinguish between "user is painting a mask" and "user is zooming in." Getting that right on iOS took more iterations than I'd like to admit.
For the AI pipeline, the server crops only the masked region (with context padding), processes it at 512x512 via LaMa, then feather-blends the result back into the full-resolution original. Non-masked pixels are completely untouched - zero quality loss outside the edit area.
The standard tools genuinely make no network requests. Easiest way to verify: open any standard tool, turn on airplane mode, and it still works. That's also why the PWA has offline support for those tools. Both img.tara.vision and our PDF toolkit (pdf.tara.vision) share about 90% of their frontend code. Same auth, billing, PWA infrastructure, UI components. Building the second product took a fraction of the time because of that.
Happy to answer questions about the architecture, privacy approach, or anything else. Feedback on the mobile UX is especially welcome - that's where I've spent most of my time lately.
Standard tools (Compress, Convert, Resize, Crop & Rotate, Strip Metadata) run entirely in your browser using Canvas API. Open your network tab - zero requests. Your files never leave your device.
AI tools (Remove Background, Remove Objects) run on our own server in Germany. Images are processed in memory and deleted immediately - no storage, no logs, no third parties. We use rembg + u2net for background removal and LaMa for inpainting, running in a Python FastAPI sidecar.
The stack: Vue 3 SPA + Laravel 12 API + PostgreSQL + Python service. PWA with offline support for the standard tools.
Why I built this: Every time I needed to remove a background or compress an image, I'd end up on some site that uploads my files to who-knows-where. For personal photos that's annoying. For client work or medical/legal documents, it's a real liability. I wanted tools I'd trust with my own data.
What's free: All 7 tools, 3 operations/day, no account needed. Pro is /month for 60/day.
Honest limitations: - The AI models are good but not state-of-the-art (tradeoff for self-hosting on modest hardware) - 7 tools vs. competitors with 50+. I'm focused on doing fewer things well. - Single founder, single server. Not enterprise-scale yet.
I'd love feedback on the UX, especially the Remove Objects tool (brush-based mask editor with pinch-to-zoom on mobile). Available in EN, DE, FR, ES, IT, MK.
Try it: https://img.tara.vision
[dead]