Rethinking Camouflage Image Generation towards a Training-Free Paradigm
The authors introduced FreeCam, a training-free method for camouflage image generation that aims to preserve a target while minimizing its visual separability from surroundings. Instead of focusing on making the target stand out, FreeCam uses inference-time steering based on semantic compatibility and appearance assimilation to synthesize realistic camouflaged images using frozen diffusion models. This enables high-quality results without needing task-specific training data.