Image generation / Model experiment
Image 1
Extend an image beyond its borders and compare candidate completions from a diffusion model.
- My contribution
- I built the image-model experiment and the candidate-selection and replay interfaces documented in the research page.
- Tools & methods
- Diffusion · U-Net · Masked DDIM
The problem
What this project investigates
Extending an image needs both plausible new content and continuity with the original. Candidate selection makes the variation between generated completions visible.
The build
What I built
The documented model uses masked DDIM sampling for outpainting and object removal. The portfolio includes real output images, a stored expansion replay and a separate studio connected through a server-side inference proxy.
Technical choices
Decisions behind the build
Preserve unmasked content
The sampling workflow targets the masked region while using the existing image as context.
Compare candidates
Batches expose alternative completions instead of presenting one output as consistently successful.
Current status
What the evidence shows
Recorded outputs and evaluation images are available. The remote inference service was not exercised in this review. A stored replay should not be described as live sampling.
Further work & evidence gaps
The write-up proposes browser inference as future work. Broader image-quality evaluation remains a useful addition.