PixelSmile: Toward Fine-Grained Facial Expression Editing

PixelSmile is a diffusion-based framework for fine-grained facial expression editing, addressing semantic overlap between expressions like fear-surprise and anger-disgust.

The authors construct the Flex Facial Expression (FFE) dataset with 60,000 images (real and anime) annotated with continuous 12-dimensional affective scores, and establish FFE-Bench to evaluate structural confusion, editing accuracy, linear controllability, and identity preservation.

PixelSmile uses flow-matching-based textual latent interpolation for continuous intensity control and a fully symmetric joint training with contrastive loss to disentangle overlapping expressions, plus an identity loss.

Experiments show PixelSmile achieves the lowest structural confusion (mSCR 0.0550) and highest editing accuracy (Acc-6 0.8627) among general editing models, and best linear control (CLS-6 0.8078, HES 0.4723) among control models, while maintaining identity similarity in the natural 0.6–0.7 range.

Ablations confirm the necessity of symmetric training, contrastive loss, and identity loss.

User study ranks PixelSmile highest in continuity (4.48) and strong identity (3.80).

The learned manifold supports expression blending, with 9 of 15 pairwise combinations producing plausible compound expressions.

The work establishes a standardized framework for continuous, controllable, and fine-grained expression editing across real and anime domains.