Reproduce “Differential Diffusion: Giving Each Pixel Its Strength” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Eran Levin, Ohad Fried
Differential Diffusion presents a framework for image editing with diffusion models, enabling granular, per-pixel control over the amount of change using a continuous "change map" rather than global strength parameters or binary masks. This approach yields more natural, spatially varied transformations and improved soft-inpainting, quantitatively demonstrating superior adherence to specified changes and receiving higher user preference for visual quality.
Key insights
- Varying the starting timestep of the diffusion process effectively controls the overall edit strength, allowing a flexible approach to global change.
- Intermediate latent representations can be selectively overwritten with noised versions of the original image to preserve fidelity in specific regions without disrupting the diffusion process.
- The spatial coherence between input images and their latent representations allows for direct, pixel-aligned translation of a desired change map into the latent space.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.