Face restoration has made steady progress, but extreme degradation still pushes models to invent details or lose the subject's identity. The authors of HonestFace argue that a more "honest" approach should reconstruct the original characteristics of the face rather than merely producing a plausible-looking result. Their method is built on a one-step diffusion model, which keeps the process efficient while targeting both identity consistency and realistic textures.
To achieve this, HonestFace uses an identity embedder that pulls identity cues from both the degraded input and several reference images of the same person. A masked face alignment module then refines fine-grained details and textural authenticity. The authors also introduce MultiRefCeleb-Test, a new real-world dataset containing multiple high-quality references paired with low-quality inputs, intended to support more realistic evaluation.
In experiments, HonestFace outperformed existing state-of-the-art methods on both visual quality and quantitative metrics. The paper, accepted at ACM MM 2026, makes its code and pre-trained models publicly available. Since this is the only source reviewed here, no independent comparison or external validation is included.