Diffusion sampling drops the single global guidance scale: per-token allocation lifts VBench total by up to 0.75
Synopsis
Classifier-free guidance for flow-matching models has applied one shared scale to all latent tokens, even though their velocity predictions evolve differently during sampling. TrajCFG maps historical changes in guided velocity direction to bounded per-token scales, giving stronger guidance to larger directional changes and weaker guidance to more consistent directions, and fuses adjacent-step with interval-normalized longer-span feedback without extra model evaluations. Across text-to-video, image-to-video, text-to-audio-video, text-to-image and image-to-3D generation it improves on base samplers; on text-to-video it raises VBench Total by 0.30 to 0.75 points over the strongest evaluated baseline on each of three backbones.
Interpretation
Guidance strength moves from one shared scale for the whole sample to a separate scale for each latent token, allocated by that token's historical change in guided velocity direction during sampling. Classifier-free guidance previously applied a single shared scale to all latent tokens, although their velocity predictions evolve differently during sampling. Controlled diagnostics and ablations examine the roles of directional feedback, token-wise allocation and history-span fusion separately.
The mechanism maps the size of the change in guided velocity direction onto a bounded scale, so larger changes receive stronger guidance and more consistent directions receive weaker guidance. It combines adjacent-step feedback with interval-normalized longer-span feedback, and reuses the predictions that standard classifier-free guidance already produces. Reusing existing predictions means no additional model evaluations are needed, so the token-wise guidance is training-free and can be added to a flow-matching sampler.
On text-to-video generation, VBench Total improves by 0.30 to 0.75 points over the strongest evaluated baseline on each of three backbones. Gains appear across text-to-video, image-to-video, text-to-audio-video, text-to-image and image-to-3D generation, spanning multiple backbones. These results support historical directional feedback as a broadly applicable signal for fine-grained guidance allocation.
Perspective
The report covers text-to-video, image-to-video, text-to-audio-video, text-to-image and image-to-3D generation, with comparisons across multiple backbones; the quantitative text-to-video result is a 0.30 to 0.75 point VBench Total gain over the strongest evaluated baseline on each of three backbones. It is aimed at researchers and engineers working on diffusion and flow-matching sampling, guidance strategies and generation-quality evaluation.
Historical directional feedback is described as a broadly applicable signal for fine-grained guidance allocation, a statement drawn from the experiments on the tasks and backbones above; how much token-wise allocation and history-span fusion each contribute is addressed by controlled diagnostics and ablations rather than by a single overall conclusion.