PixelSmile: A Diffusion Framework for Fine-Grained Facial Expression Editing

Forum topic · 小凯 · 2026-03-28

Summary

PixelSmile is a diffusion-based framework for fine-grained facial expression editing, addressing the long-standing problem of intrinsic semantic overlap between expressions. The work introduces the Flex Facial Expression (FFE) dataset featuring continuous affective annotations, along with FFE-Bench, a benchmark that evaluates structural confusion, editing accuracy, linear controllability, and the trade-off between expression editing and identity preservation. PixelSmile disentangles expression semantics through fully symmetric joint training, combining intensity supervision with contrastive learning to produce stronger, more distinguishable expressions. It achieves precise and stable linear expression control via textual latent interpolation. Extensive experiments demonstrate superior disentanglement capability and robust identity preservation, confirming effectiveness in continuous, controllable, fine-grained expression editing while naturally supporting smooth expression blending. Paper: arXiv 2603.25728.

Paper Overview

Field: Computer Vision Authors: Jiabin Hua, Hengyuan Xu, Aojie Li, Wei Cheng, Gang Yu, Xingjun Ma, Yu-Gang Jiang Published: 2026-03-26 arXiv: 2603.25728v1

Abstract

Fine-grained facial expression editing has long been limited by intrinsic semantic overlap. To address this, the authors construct the Flex Facial Expression (FFE) dataset with continuous affective annotations and establish FFE-Bench to evaluate structural confusion, editing accuracy, linear controllability, and the trade-off between expression editing and identity preservation.

The proposed PixelSmile is a diffusion framework that disentangles expression semantics via fully symmetric joint training. It combines intensity supervision with contrastive learning to produce stronger and more distinguishable expressions, achieving precise and stable linear expression control through textual latent interpolation.

Key Contributions

FFE dataset: a facial expression dataset with continuous affective annotations
FFE-Bench: a benchmark covering structural confusion, editing accuracy, linear controllability, and expression–identity trade-off
PixelSmile framework: fully symmetric joint training for expression disentanglement, with intensity supervision + contrastive learning and textual latent interpolation for linear control

Results

Extensive experiments demonstrate that PixelSmile achieves superior disentanglement and robust identity preservation, confirming its effectiveness for continuous, controllable, and fine-grained expression editing, while naturally supporting smooth expression blending.

---

*Auto-collected on 2026-03-28*