Diffusion Guidance
A technique that injects additional gradients or biases into the denoising network during the sampling stage of a diffusion model, steering generated samples toward a target distribution without retraining the model.
Timeline (last 90 days)
At sampling time, diffusion guidance shifts each emulated climate model distribution toward observations, with its strength selected through perfect-model experiments.
On seventeen synthetic targets, guidance more than halves the median bias of the 1st and 99th percentiles of daily temperature, improving probability scores and reliability.
Under three emissions scenarios, guidance reduces the median RMSE of global mean temperature by 32-57%.
In detailed benchmarks, guidance outperforms existing bias correction methods on daily field probability scores.
Diffusion guidance injects additional gradients or biases into the denoising network during the sampling stage to steer generated samples toward a target distribution without retraining the model.
All Facts (5)
At sampling time, diffusion guidance shifts each emulated climate model distribution toward observations, with its strength selected through perfect-model experiments.
50%UnverifiedOn seventeen synthetic targets, guidance more than halves the median bias of the 1st and 99th percentiles of daily temperature, improving probability scores and reliability.
50%UnverifiedUnder three emissions scenarios, guidance reduces the median RMSE of global mean temperature by 32-57%.
50%UnverifiedIn detailed benchmarks, guidance outperforms existing bias correction methods on daily field probability scores.
50%UnverifiedDiffusion guidance injects additional gradients or biases into the denoising network during the sampling stage to steer generated samples toward a target distribution without retraining the model.
50%