Diffusion models generate images through iterative denoising, but this process can accumulate errors. The root cause is exposure bias: during training, the denoiser is fed real images with added noise, but at inference it receives its own imperfect predictions. This mismatch can lead to degraded samples over many steps.
The paper introduces correction training, a method designed to align training and inference conditions. While the abstract does not detail the mechanism, the goal is to reduce error accumulation and improve sampling fidelity.
If effective, correction training could make diffusion models more reliable for tasks requiring many sampling steps, such as high-quality image generation. The work highlights a known weakness in diffusion models and offers a potential fix.