Preprint · 2026 Training-free guidance

TrajCFG: Token-Wise Classifier-Free Guidance via Denoising Trajectory Feedback

Let every token follow the guidance it needs—using signals already present in its denoising history.

Jingyu Liu1,†,§ Hao Jiang2,† Yikang Ding2,‡ Huafeng Shi2 Xu Yang2 Xiaoqiang Liu2 Guoxin Zhang2 Pengfei Wan2 Kun Gai2,*

1Renmin University of China2Kling Team, Kuaishou Technology

† Equal contribution§ Work done at Kling Team, Kuaishou Technology‡ Project Leader* Corresponding author

Paper Code
Overview of TrajCFG: token trajectories provide feedback that adaptively changes guidance scales during denoising.
At a glance Historical token trajectories become guidance feedback.
0training steps
Token-wiseadaptive scales
0 extramodel evaluations
5generation modalities
01 Overview

TL;DR

One shared CFG scale overlooks how differently tokens evolve.

TrajCFG turns denoising history into a fine-grained control signal.

Classifier-free guidance typically applies a shared scale to all latent tokens, although their velocity predictions evolve differently during sampling. TrajCFG adapts guidance token by token using historical changes in guided velocity direction: larger directional changes receive stronger guidance, while more consistent trajectories receive weaker guidance.

The method combines adjacent-step and interval-normalized longer-span feedback, reuses standard CFG predictions, and requires neither training nor additional model evaluations. Across text-to-video, image-to-video, text-to-audio-video, text-to-image, and image-to-3D generation, it consistently improves generation quality on multiple backbones.

A

Read the trajectory

Measure token-level directional changes already produced along the denoising path.

B

Allocate guidance

Map trajectory feedback to bounded token-wise scales instead of one global value.

C

Plug in and sample

Reuse standard conditional and unconditional predictions with no added NFE.

02 Method

Trajectory feedback

Guidance that listens to each token’s recent motion.

As sampling unfolds, tokens reveal whether their velocity direction is stable or still changing. TrajCFG converts that signal into a bounded, bidirectional adjustment around the base guidance scale.

TrajCFG method pipeline showing standard classifier-free guidance predictions, trajectory feedback and token-wise scale adaptation.
Figure 1. TrajCFG reuses the standard CFG branch outputs and adapts scales from denoising history.
Plot relating historical directional angle to token content relevance.
Controlled observation. Historical directional changes expose content-relevant token behavior.
  1. 01

    Compare velocity directions

    Track adjacent-step change and a normalized longer-span change for every latent token.

  2. 02

    Fuse short and long history

    Combine responsive local feedback with a steadier interval signal to reduce noisy allocation.

  3. 03

    Adapt within safe bounds

    Convert the fused feedback into token-wise CFG scales while preserving the original sampler structure.

03 Results

Quality without overhead

Consistent gains, at the runtime of standard CFG.

+0.30–0.75VBench Total over the strongest evaluated baseline across three T2V backbones
1.0×relative runtime, versus 1.5× for additional-evaluation guidance methods
Text-to-video
Text-to-video quantitative comparison across multiple backbones and guidance baselines.
TrajCFG improves overall VBench performance across all evaluated video backbones.
Image-to-video
Image-to-video quantitative comparison across backbones and guidance baselines.
The same token-wise feedback generalizes to image-conditioned video generation.
05 Citation

Build on TrajCFG

If this work helps your research, please cite it.

Preprint available on arXiv: 2610.14153.

BibTeX
@article{liu2026trajcfg,
  title   = {TrajCFG: Token-Wise Classifier-Free Guidance
             via Denoising Trajectory Feedback},
  author  = {Liu, Jingyu and Jiang, Hao and Ding, Yikang and
             Shi, Huafeng and Yang, Xu and Liu, Xiaoqiang and
             Zhang, Guoxin and Wan, Pengfei and Gai, Kun},
  journal = {arXiv preprint arXiv:2610.14153},
  year    = {2026},
  eprint  = {2610.14153},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url     = {https://arxiv.org/abs/2610.14153}
}
BibTeX copied