PixelSmile addresses fine-grained facial expression editing in images, offering continuous control, reduced semantic entanglement, and strong identity preservation. It targets researchers and power users seeking precise and realistic manipulation of facial expressions while maintaining subject integrity. The project provides a novel approach to expression editing with enhanced control and fidelity.

How It Works

The system builds upon the Qwen/Qwen-Image-Edit-2511 base model, enhanced with PixelSmile LoRA weights. This architecture enables fine-grained control over expressions, minimizing unintended semantic changes and preserving the original identity of the face. The approach aims for a more nuanced and controllable editing process compared to existing methods.

Quick Start & Requirements

Installation involves cloning the repository, creating a Python 3.10 conda environment, and running pip install -r requirements.txt. A critical step is patching the diffusers installation using bash scripts/patch_qwen_diffusers.sh. Base models (Qwen/Qwen-Image-Edit-2511) and PixelSmile-preview weights are available for download via Hugging Face. A live online demo is accessible, and an arXiv paper is published. Community support for ComfyUI is available via a separate repository.

Highlighted Details

• Features fine-grained facial expression editing with continuous control.
• Offers reduced semantic entanglement and strong identity preservation.
• Provides a live online demo and an arXiv paper.
• Includes inference code and benchmark data.
• Community support for ComfyUI is available.

Maintenance & Community

Community contributions include a ComfyUI implementation. No explicit details on active community channels (e.g., Discord, Slack), sponsorships, or partnerships are provided in the README.

Licensing & Compatibility

The repository's license is not specified in the README, which is a critical omission for assessing compatibility, especially for commercial use or integration into closed-source projects.

Limitations & Caveats

Currently, only "Preview" model weights are available, with stable versions and training code slated for future release. The project is in active development, and features like stable weights, training code, and benchmark code are still pending.