SceneLock

• 

  picture of prepared tshirt with plain backdrop

• 

  website workflow

• 

  picture of before and after (image addition)

Inspiration

I wanted to explore how small sellers could create better product photos using images they already have. Setting up a new scene for every photo can be difficult, so I started with a simple question: could someone take a basic product photo and turn it into something they would want to use in a listing?

That became SceneLock: a tool for placing a product into a new scene, with enough editing controls to make the result feel more natural.

What it does

Users upload a product photo, remove its background, and fix any parts of the cutout that need attention. They can then upload a background or generate one using FLUX.1-schnell.

From there, they can resize, rotate, and move the product, adjust colour matching and shadows, and download the finished image. SceneLock also lets users remove small distractions from a background and download that cleaned image separately.

How we built it

I used Python and Streamlit for the app, rembg with a pretrained U2Net model for background removal, and Pillow, NumPy, and SciPy for image processing. Background cleanup uses OpenCV inpainting, while generated backgrounds come through a hosted FLUX.1-schnell demo.

ChatGPT helped with coding and troubleshooting. The initial prototype existed before the hackathon; later work included compatibility and setup fixes.

I developed the workflow around testing actual images: upload a product, try a background, inspect the result, and work out which controls would help improve it.

Challenges we ran into

Getting the product onto a background was one thing. Making it look like it belonged there was harder. Some early results looked too polished or disconnected from the scene, which led me to add matching adjustments and shadows.

Cutout edges sometimes needed manual fixes. Background generation had usage limits, and removing objects from detailed backgrounds could leave smudges. I also ran into Python dependency and Windows security-policy issues while getting the app running.

Accomplishments that we're proud of

I’m proud that the app now supports a complete workflow, from uploading a photo to downloading an edited result. The manual controls, comparisons, and undo options give users room to experiment. Standalone background cleanup also became a useful feature of its own.

What we learned

I learned how much small details matter in image editing. A rough edge or mismatched shadow can make the whole scene feel wrong. I also learned that a useful AI app needs ways for people to inspect and correct its output.

What's next for SceneLock

I’d like to improve difficult cutouts, object cleanup, and lighting matching, and make background generation more dependable. Most importantly, I want to test it with small sellers and learn what would make it useful in their everyday work.

Built With

• flux
• scipy
• u2net