Summary

WithAnyone addresses controllable and identity-consistent image generation, tackling "copy-paste" artifacts in face synthesis. It breaks the trade-off between facial similarity and controllability of expressions, hairstyles, or poses. The project enables generating multiple distinct identities harmoniously within a single image, targeting researchers and power users for advanced face manipulation and group scenarios.

How It Works

The core approach balances SigLIP embeddings (for semantic details/stylization) with ArcFace embeddings (for identity preservation) via a "Resemblance in Spirit" vs. "Resemblance in Form" slider. This allows fine-grained control over expressions, accessories, and hairstyles while maintaining consistent identity. The architecture also supports generating multiple identities that integrate naturally into a single image.

Quick Start & Requirements

Installation uses pip install -r requirements.txt. Significant checkpoint downloads (~51 GB) are required, including black-forest-labs/FLUX.1-dev. Inference necessitates face bounding boxes, provided manually or automatically. A HuggingFace Space demo and Gradio applications (gradio_app.py, gradio_edit.py) are available for testing. Links to models and demos are on HuggingFace.

Highlighted Details

• Mitigates "copy-paste" artifacts and resolves the similarity-controllability trade-off in face generation.
• Enables multi-identity image generation for harmonious group compositions.
• Features a slider for granular control over identity preservation vs. stylistic freedom.
• Supports LoRA checkpoints for enhanced stylization.
• Offers specialized models for text-to-image and face editing using FLUX.1 Kontext.

Maintenance & Community

Recent updates (Oct 2025) include demos, model checkpoints, and datasets. Training codebase release is planned post-1k stars. No explicit community channels or roadmaps are detailed.

Licensing & Compatibility

Code is Apache 2.0. Models and datasets are strictly for non-commercial academic research under the "FLUX.1 [dev] Non-Commercial License v1.1.1". Commercial use, profit redistribution, or violations of law/ethics are prohibited. Users must comply with all applicable licenses.

Limitations & Caveats

Preliminary versions (.K.preview, .Ke.preview) may have stability/quality issues. ArcFace model setup might require manual intervention. Inference requires explicit face bounding boxes. Effective control depends on detailed prompts.