The field of diffusion models is moving towards improving the balance between prior contribution and data fidelity term, with a focus on developing adaptive guidance strategies and merging pretrained experts to break the likelihood-quality trade-off. Innovative approaches, such as reweighted losses and physics-guided control nets, are being explored to advance the field. Noteworthy papers include:

• A paper that proposes an adaptive likelihood step-size strategy, resulting in improved reconstruction quality across diverse imaging tasks.
• A paper that introduces a simple plug-and-play sampling method, combining two pretrained diffusion experts to match or outperform its base components.
• A paper that derives a new theoretical interpretation of reweighted losses, resulting in reduced data-model KL-divergences and improved sample quality.
• A paper that proposes a blind adaptive local denoising method for CEST imaging, exploiting the self-similar nature of CEST data to derive an adaptive variance-stabilizing transform.
• A paper that proposes a physics-guided control net for generative spatially varying image deblurring, reconciling model-based and generative paradigms to achieve superior perceptual quality and physical accuracy.