Classifier-free guidance (CFG) is a common technique for steering diffusion models toward desired outputs, but it typically applies a single shared scale to every latent token. That uniform treatment may be suboptimal, because the velocity predictions for different tokens evolve at different rates during the sampling process.
TrajCFG, introduced in a new arXiv paper, proposes a token-wise alternative. Instead of a fixed global scale, it uses feedback from the denoising trajectory to adjust guidance for each token individually. The method is designed to account for the fact that some tokens need stronger or weaker guidance at different points in sampling.
The available abstract is truncated, so the paper's experimental setup and quantitative results are not described in the source. The core idea, however, is clear: move from one-size-fits-all guidance to a trajectory-adaptive, token-specific approach.