Researchers introduced PEEK, a training-free plugin for diffusion image generators. It makes an extra denoiser evaluation during selected sampling steps, then uses that look-ahead prediction to adjust the next step and reduce exposure bias—the mismatch between training inputs and the imperfect inputs a model encounters while generating.
The authors report improvements over baselines matched for denoiser evaluations across eight pixel- and latent-space models, including text-guided setups. They report FID improvements of up to 49.3%; calibrating the plugin for a model took at most 2.2 minutes in their experiments.
Why it matters
For teams researching diffusion image generators, PEEK offers a sampler-side method to test without retraining, though the paper does not establish how it performs in deployed products.
