phrame

Adobe Studio, lightroom, and numerous pro image editing software forces small creators and photographers to choose between professional tools that are difficult to learn and AI editors that remove too much. This is why we built the project Phrame to bridge that gap: our goal is to let AI handel the difficult and repetitive tasks alongside the creator.

We split phrame into three main layers:

The editing workspace that handles the canvas, imported images, controls, and traditional editing tools. The application backend that validates requests, protects credentials, and connects the editor to AI services. Finally, the inference layer that processes resource-intensive image generation and editing tasks.

AI requests go through a server-side proxy deployed with Vercel, which keeps model credentials out of the browser. We also designed a desktop backend for local workflows and made the inference system flexible enough to support different models and GPU providers.

We developed features through separate branches and pull requests so work on the interface, editing tools, AI integration, and GPU infrastructure could happen with fewer conflicts.

One challenge was combining predictable manual tools with AI outputs that can change between generations. We needed to make it clear when an edit would be exact, when AI was involved, and when the application was still processing.

Performance was another issue. High-resolution images use a lot of memory, and diffusion-based editing requires much more computing power than a normal browser tool. Separating browser interactions from GPU processing helped keep the editor responsive.

We also worked through credential security, long requests, unavailable models, deployment limits, and merge conflicts. These problems taught us to separate responsibilities more clearly and keep each pull request focused.

We learned that building an AI product takes much more than connecting an interface to a model. Request validation, state management, security, error handling, latency, and clear user feedback are just as important as the model itself.

We also found that AI works best alongside traditional tools. Manual controls give users precision, while AI makes complex edits faster.

Next, we plan to complete the NVIDIA inpainting pipeline, improve masking, selections, layers, and edit history, and add progress indicators, cancellation, comparison views, and nondestructive AI editing.

Our long-term goal is to make phrame an accessible creative workspace where manual control and generative AI work together.