ArtQuilt

• 

  before

• 

  after

Inspiration

I wanted to explore how a single photograph can be reimagined as a small gallery: different corners of the same image interpreted by different artistic voices. The idea of stitching together contrasting art styles into one coherent image — like quilting with paintings — felt playful, surprising, and visually expressive.

What it does

ArtQuilt transforms a photo into a single composite artwork by splitting the image into organically shaped regions and restyling each region with a different artistic flavor (for example, a cubist corner, an abstract block, an impressionist wash, and a surreal vignette). The pieces are blended back together so the final image reads as a unified, eclectic canvas.

How I built it

I started with a small Python command-line tool that generates soft, wavy region masks and extracts region crops. Each region is sent to an image-to-image model with a distinct prompt to produce a stylistic transformation; the returned images are composited using alpha masks so seams are soft. The project uses established image libraries for masks and compositing and a remote inference endpoint to keep iteration fast during the hack.

Challenges I ran into

• Getting the region boundaries to feel organic rather than obvious straight cuts required iterating on noise and smoothing techniques.
• Balancing how much of a region the model should change: heavy stylization can lose the subject, light stylization can feel like a filter. Finding useful prompts took several experiments.
• Handling partial failures and edge cases (small or empty regions, model timeouts) without breaking the whole image composite.

Accomplishments that I am proud of

• The visual concept works: compositions produce surprising, gallery-like results where each quarter can tell a different visual story while the whole still reads as a single artwork.
• Implemented smooth, randomized region boundaries and soft alpha blending, which greatly improves perceived continuity across style transitions.
• Built a lightweight CLI to experiment with deterministic or randomized behavior and to limit how many remote calls to make during development.

What I learned

• Prompt engineering matters — small changes to style prompts drastically change results and how recognizably the original image survives the transformation.
• Soft masks and compositing are as important as the stylization: without them, seams look jarring and the effect feels like a collage rather than a unified piece.
• Iterating quickly with a remote inference API accelerates experimentation, but you need to design for quota/cost and occasional failures.

What's next for ArtQuilt

• Improve style choreography: let users pick or upload style references for each region (or sample from a style gallery).
• Implement a lightweight local inference fallback so users can run transformations offline, and add GPU support for speed.
• Add a small web UI for interactive region selection and re-styling previews, plus presets and downloadable high-res outputs.
• Experiment with more than four regions or non-grid compositions, and add a simple color-harmonization pass so transitions feel even more cohesive.

Built With

• api
• huggingface
• python