Curious about AI, I transitioned from mere experimentation to creating a real product—Photfix, my AI-powered app that enhances low-quality images. Users can upload blurry photos, let AI upscale them to 2K or 4K, and download the results. Initially, I faced challenges with APIs like Replicate causing slow latency, leading me to explore open-source models such as Real-ESRGAN and self-hosting via GPU containers on Modal.com. My breakthrough came when optimizing Docker containers by embedding model weights directly, reducing cold start times from 30 to under 5 seconds. I discovered that GPU types significantly impact performance; the NVIDIA L40S proved superior for image tasks. Now live, Photfix efficiently upscales profile pictures and restores old photos. Key lessons include the importance of controlling latency, pre-baking weights, and testing GPUs. Looking ahead, I plan to introduce batch processing and face-aware enhancements. Check it out at Photfix!
Share
Read more