The field of image generation and steganography is witnessing significant developments, with a focus on improving the security, capacity, and perceptual quality of visual content. Recent innovations have leveraged generative models, particularly diffusion models, to enable controllable information embedding and adaptive image synthesis. Notable advancements include the integration of bit-position locking and diffusion sampling injection to achieve a favorable balance between randomness and constraint, enhancing robustness against steganalysis without compromising image fidelity. Moreover, training-free watermarking frameworks have been proposed for autoregressive image generation models, utilizing the redundancy property of codebooks to embed watermarks without affecting image quality. The refinement of autoregressive image generation models has also been explored, with new paradigms enabling iterative refinement of previously generated content. Furthermore, unified frameworks for conditional panoramic image generation have been developed, addressing the limitations of existing approaches and integrating text and image conditioning into cohesive architectures. Lastly, the generalization mechanisms of diffusion probabilistic models have been revisited, revealing that generalization in natural data domains is progressively achieved during training before the onset of memorization, and that principled early-stopping criteria can optimize generalization while avoiding memorization. Noteworthy papers include:
- Shackled Dancing: A Bit-Locked Diffusion Algorithm for Lossless and Controllable Image Steganography, which introduces a plug-and-play generative steganography method.
- Training-Free Watermarking for Autoregressive Image Generation, which proposes a simple yet effective match-then-replace method for watermarking.
- TensorAR: Refinement is All You Need in Autoregressive Image Generation, which enables iterative refinement of previously generated content.
- Conditional Panoramic Image Generation via Masked Autoregressive Modeling, which integrates text and image conditioning into a cohesive architecture.
- Bigger Isn't Always Memorizing: Early Stopping Overparameterized Diffusion Models, which reveals the generalization mechanisms of diffusion probabilistic models.