Training-Free Self-Correction for Multimodal Masked Diffusion Models
This work addresses a key bottleneck in multimodal generation for AI practitioners by offering a robust, incremental improvement without additional training.
The paper tackles error accumulation in masked diffusion models during sampling by proposing a training-free self-correction framework that exploits pre-trained model biases, resulting in significant improvements in generation quality on text-to-image and multimodal tasks with reduced sampling steps.
Masked diffusion models have emerged as a powerful framework for text and multimodal generation. However, their sampling procedure updates multiple tokens simultaneously and treats generated tokens as immutable, which may lead to error accumulation when early mistakes cannot be revised. In this work, we revisit existing self-correction methods and identify limitations stemming from additional training requirements or reliance on misaligned likelihood estimates. We propose a training-free self-correction framework that exploits the inductive biases of pre-trained masked diffusion models. Without modifying model parameters or introducing auxiliary evaluators, our method significantly improves generation quality on text-to-image generation and multimodal understanding tasks with reduced sampling steps. Moreover, the proposed framework generalizes across different masked diffusion architectures, highlighting its robustness and practical applicability. Code can be found in https://github.com/huge123/FreeCorrection.