I built ClipLumi around a problem I kept running into with generative image tools: **generation and editing are not the same job**. When you are exploring from a blank canvas, regeneration is great. You can ask for ten different directions and accept that every result may reinterpret the scene. But once you already have an image that is 80–90% correct, that same behavior becomes expensive. You may only want a cleaner background, one removed object, a repaired old photo, or a small change to a product shot. A full regeneration can fix the requested part and quietly damage three things you wanted to keep. That led to the design principle behind ClipLumi: **change the requested region, preserve the rest of the useful visual**. The practical workflow is deliberately simple: 1. Start from an image you already want to keep. 2. Describe the specific change in plain language. 3. State important preservation constraints when they matter — for example, keep the product label, shape, colors, reflections, camera angle, face, clothing, or composition unchanged. 4. Generate the edit. 5. Compare the result with the original, not just with the prompt. 6. Make another focused pass only if the next change solves a real problem. That sounds obvious, but it changes how I think about AI image quality. A result is not successful just because the edited area looks good. It also has to avoid unnecessary drift outside the requested change. For a product image, “replace the background with a clean studio setting” should not also redesign the product or distort printed text. For a portrait, “remove the person in the background” should not alter the main subject’s face or clothing. For restoration, repairing scratches and fading should not silently change who the people in the photograph look like. ClipLumi is browser-based because I wanted the edit loop to stay lightweight. Many people who need these changes are not trying to build a complex layered composition. They are marketers preparing an asset, e-commerce operators cleaning a product image, creators fixing a generated visual, or someone restoring a personal photo. The job is often “make this one correction, then let me decide whether it is good enough.” The current product supports natural-language image editing and reference-guided workflows where the selected model supports them. I am intentionally not claiming that every model can perform every kind of edit or that AI replaces professional retouching. Pixel-perfect compositing, print color management, and deeply layered production work can still belong in dedicated graphics tools. The part I am most interested in is the review loop. I think AI image editors need better ways to help users reason about **what changed unintentionally**, not only what changed successfully. In other words, the future of image editing is probably not just “better prompts” or “bigger models”; it is also better preservation, comparison, and repair workflows. If you try ClipLumi, I would especially value feedback on one question: **when you ask for one focused edit, does the result preserve the parts of the image you actually wanted to keep?** You can try it at https://cliplumi.com/.

ClipLumi

Edit existing images with natural-language instructions

Edit existing images with natural-language instructions