Vivu — AI Photo Editing for Travel Memories

The challenge

Travel photos often capture more than the moment people want to remember: strangers in the background, signs, litter, or other visual distractions. Removing them traditionally requires editing experience, even though the desired outcome is simple to describe.

Vivu turns that task into a guided three-step experience: upload a photo, select what should disappear, and receive a completed image. The product had to make a capable AI workflow feel approachable while preserving enough control for people to correct a selection, compare results, and continue refining the same photo.

What I built

I developed Vivu as an end-to-end AI product rather than a model demonstration. My work covered the editing experience, the image-processing workflow, the production operation around the models, and the supporting product systems that let people try the service, return to previous work, and share it with others.

The same core journey was designed for the public browser experience and a native iPhone experience: choose an image, prepare it for editing, mark the unwanted area, wait for processing, then compare, save, or share the result.

AI-assisted editing

Vivu combines two complementary forms of computer vision:

• Generative image inpainting reconstructs the area behind a removed person or object so that the result blends with the surrounding scene.
• Visual segmentation supports intelligent object selection, turning a click into an editable region instead of asking the user to trace every boundary.
• Mask-based editing keeps the user in control. An intelligent selection can be combined with manual cover and uncover tools, then adjusted with undo and redo.

The models remain behind the product experience. People do not need to choose a model or understand its parameters; they decide what to remove and can refine the result when the first edit is not enough.

Product experience

The editor supports the complete workflow around the AI result:

1. Start immediately as a guest or sign in to keep a longer-lived workspace.
2. Select an object intelligently or paint the region manually.
3. Pan and zoom without losing the alignment between the photo and its edit mask.
4. Compare the original and edited image, then continue editing the latest result.
5. Download or share the finished photo.
6. Return to a photo library and inspect previous editing steps when signed in.

First-run guidance introduces the workflow inside the editor, while clear waiting, failure, and recovery states keep the user oriented during longer AI operations. Referral and usage experiences provide a lightweight way for the product to grow without interrupting the central editing task.

From AI workflow to production

Shipping Vivu meant designing for conditions that a local model demo rarely has to handle. Image processing runs asynchronously so the editing interface can preserve the user’s work while a request is queued or processed. Each operation moves through explicit states, stores its inputs and outputs durably, and can recover from an interrupted dispatch or a stalled task.

The AI runtime manages model resources independently from the user-facing product. It limits conflicting work, releases expensive resources when idle, and keeps short-lived selection context only as long as it remains useful. Retries, ownership checks, usage limits, and controlled feature rollout protect both the user workflow and the service under real production load.

These boundaries allowed the editing experience and the AI processing layer to evolve independently without exposing infrastructure choices to the user.

Product in the wild

Vivu is publicly available at vivu.study. Its public presence continues beyond the editor: the Vivu YouTube channel shares visual demonstrations, while the Vivu Facebook page provides a place for product updates, user education, and community discovery.

Together, the product site and social channels connect discovery, learning, and hands-on use instead of treating distribution as separate from the product.

Outcome

Vivu moved from an image-processing workflow to a public AI product with a coherent editing journey, production safeguards, persistent user work, guided onboarding, and community-facing channels. The result demonstrates how a specialized model capability can become useful when control, recovery, and product communication are designed around it.

What I learned

Model quality is only one part of an AI editing product. Selection ergonomics, visible progress, recoverable state, and a clear path from first visit to finished image have just as much influence on whether the experience feels trustworthy. The strongest product decisions made the AI less visible and the user’s intent more direct.