A new paper from vivo's internet technology team introduces Self-Swapping Guidance (SSG), a novel guidance paradigm for diffusion models that enhances image generation without relying on text prompts, additional training, or added noise. The core idea is simple: during the denoising process, tokens are directly swapped within the model's internal representations to steer generation toward higher-quality outputs. This approach contrasts with traditional methods that either require text conditioning, specialized training, or introduce noise that can degrade image fidelity. SSG is presented as a lightweight strategy that works with existing models, making it potentially easy to integrate into current pipelines. The paper was accepted as an Oral presentation at CVPR 2026, indicating strong peer recognition. For researchers and engineers working on generative models, SSG offers a fresh perspective on guidance mechanisms and could lead to more efficient and controllable image synthesis in practical applications.
Researchers propose Self-Swapping Guidance (SSG), a new paradigm for diffusion models that improves generation by directly swapping tokens, eliminating the need for text conditions or additional noise. This lightweight approach addresses limitations of traditional guidance methods and achieves stronger results. Accepted as a CVPR 2026 Oral, it signals a promising direction for more efficient and controllable image generation.