# A photo editor for 2030 Feather, Inter IIT Tech Meet 14.0 ยท Team Lead, Product and Tech Dec 2025 Adobe's problem statement at Inter IIT Tech Meet 14.0 asked for the photo editor of 2030, running on limited hardware. Feather was our answer: a phone editor where you describe an edit in plain words, restyle a photo after a piece of art, and get help writing the prompt as you type. I led a six-person team across research, design, ML and engineering, as the lead for product and tech. We placed 8th of 23 IITs and presented to Adobe's product team. **Feather, screen by screen** An empty project, with Main and Branch at the top and AI and Expert beside them. A photo in, and Feather suggests edits for it: "Enhance the warm, golden light". Ghost text. You type "Make it noir and reduce it's" and the phone offers the rest. Suggestions change with the photo: a portrait gets "Enhance pink hair, mute background". The Smart Adjust pad. Touch the point between bright, dark, sharp and soft. Expert mode swaps the pad for sliders: exposure, contrast and the rest. Frames from the team's screen recordings of the working app. ## Split between phone and cloud "Limited hardware" decided the architecture. Anything that has to feel instant runs on the phone, and anything heavy goes to the cloud. **What runs where** The line between them is latency. If you are waiting on every keystroke, it can't go over the network. A quantized Llama 3.2 1B runs on the phone and finishes your prompt as you type. It works offline and costs nothing per suggestion. SDXL runs in the cloud on Modal for style transfer, Gemini refines prompts and suggests edits, and filters run in real time on the phone. ## Help with the words Describing an edit is harder than it sounds. "Make it better" gives a model nothing to work with. So the prompt box finishes your sentence in grey as you type, a Refine button turns a short prompt into a detailed one, and if typing is the problem, you can talk to the assistant instead. **Ghost text in the prompt box** Feather also looks at the photo and suggests edits for it. On a sunset over the water it offered "Enhance the warm, golden light". On a portrait with pink hair, "Enhance pink hair, mute background". The suggestions are the prompt you might have written if you had known what to ask. ## When words aren't enough Some edits are easier to point at than to say. Smart Adjust has a pad with bright and dark on one axis and sharp and soft on the other. You touch the feel you want, and in the app's own words, "the more you refine, the better Kimi adapts to your unique creative style." Try it on the photo from our demo. **Pointing instead of describing** Our demo photo with a simple stand-in for the edit: brightness up and down, contrast or softness left and right. ## Room to try things Every project has a main line and branches, so you can try a direction without losing the one you had. A switch moves between AI and Expert modes, and Expert swaps the pad for real sliders, exposure, contrast and the rest, for people who'd rather set the numbers themselves. The app is React Native on Expo, with the heavy models behind it on Modal. **What I'd measure** Edits people keep - **Ghost text taken**suggestions people accept as they type - **Undo rate**edits reversed straight away - **Time to first edit**from opening a photo - **Cloud cost per kept edit**what SDXL costs for each image kept The last one is the one Adobe's brief was really about. On limited hardware, a cloud edit nobody keeps is pure cost, and the phone should be doing everything it can before you pay for the cloud. [GitHub](https://github.com/NiharRD/ArtStyleTransfer)