PDGrad: Guiding Diffusion Model for Reference-Based Blind Face Restoration with Pivot Direction Gradient Guidance
Geon Min, Tae Bok Lee, Yong Seok Heo · 2024
Read PDGrad: Guiding Diffusion Model for Reference-Based Blind Face Restoration with Pivot Direction Gradient Guidance by Geon Min, Tae Bok Lee, Yong Seok Heo (2024) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/pmc-11598124.txt.
Description
Geon Min, Tae Bok Lee, and Yong Seok Heo (2024) improve a method for repairing damaged face photos with help from a clean reference photo of the same person. Guided diffusion models can do this repair without new training by following gradients from several loss functions, each pushing the result toward a desired trait. The trouble is that these gradients can fight each other, so the repair ends up weaker than it should be.
Their method, called PDGrad, defines losses on both fine detail and high-level features, using the rough repair and the reference photo together. When gradients conflict, one gradient per level is chosen as a pivot and the others are rotated to agree with it; any adjusted gradient that grows too large is scaled back. On the CelebRef-HQ test set, the method beat rival approaches on both measured scores and visual quality.
Catalog record
| CREATOR | Geon Min, Tae Bok Lee, Yong Seok Heo |
|---|---|
| YEAR | 2024 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 43 KB |
| VERSION | 1.0 |
| LICENSE | CC BY 4.0 |
| RIGHTS | CC BY |
| SOURCE | Europe PMC |