Ammmob/PixelSmile

PixelSmile: Fine-grained facial expression editing with continuous control, reduced semantic entanglement, and strong identity preservation.

What it solves

PixelSmile enables fine-grained editing of facial expressions in images, allowing users to change the emotion of a person (or anime character) while maintaining their identity and the overall image context.

How it works

It builds upon the Qwen-Image-Edit-2511 base model, using a LoRA (Low-Rank Adaptation) approach to specialize the model for expression editing. During training, it utilizes CLIP encoders for text/image understanding and InsightFace (ArcFace) to ensure identity embeddings are preserved, preventing the person's face from changing into someone else during the edit.

Who it’s for

  • AI researchers and developers working on image manipulation and facial analysis.
  • Digital artists and creators who need precise control over facial expressions in generated or existing images.
  • Users of ComfyUI looking for advanced conditioning interpolation for facial edits.

Highlights

  • Fine-grained control: Specifically designed for precise facial expression modifications.
  • Identity preservation: Uses identity embeddings to keep the subject recognizable.
  • Multi-domain support: Supports both human and anime facial expressions.
  • Flexible deployment: Provides inference code, training scripts, and community support for ComfyUI.