Overview
Standard classifier-free guidance (CFG) applies a uniform, shared guidance scale across all latent tokens during generation. However, latent tokens exhibit distinct velocity prediction dynamics throughout the sampling trajectory. To address this discrepancy, researchers introduced Trajectory-Adaptive Classifier-Free Guidance (TrajCFG), a training-free framework designed for flow-matching generative models. TrajCFG dynamically modulates guidance on a per-token basis by monitoring historical directional changes in guided velocity predictions.
Research Context
In conventional flow-matching sampling setups, CFG enforces a constant scale across the entire token field. Because different tokens do not evolve at identical rates or follow identical vector trajectories during denoising, applying an unvarying scale can overlook fine-grained differences across latent tokens. Existing approaches typically lack token-specific calibration or require additional computation to determine how guidance should be distributed.
Approach
TrajCFG operates as a training-free formulation that extracts guidance signals directly from the denoising trajectory without requiring extra neural network evaluations. The system reuses standard CFG predictions and maps directional history into bounded token-wise scales:
- Directional Feedback Mapping: TrajCFG measures changes in guided velocity direction. Larger shifts in direction trigger stronger guidance scales, whereas more consistent trajectories receive weaker guidance.
- Multi-Span Fusion: The method combines adjacent-step feedback with interval-normalized longer-span feedback to inform guidance allocation.
- Computational Efficiency: By computing token-level scaling strictly from already generated CFG velocity estimates, TrajCFG incurs no additional model forward passes.
The researchers conducted controlled diagnostics and ablation studies to isolate the impact of its key components, specifically analyzing the contributions of directional feedback, token-wise guidance allocation, and history-span fusion.
Findings
TrajCFG was evaluated across five generative modalities: text-to-video, image-to-video, text-to-audio-video, text-to-image, and image-to-3D generation. Key empirical findings include:
- Consistent Improvements Across Samplers: TrajCFG outperformed base sampling baselines across multiple distinct model backbones.
- Text-to-Video Benchmarks: In text-to-video evaluations, TrajCFG increased VBench Total scores by 0.30 to 0.75 points compared to the strongest evaluated baseline across three separate backbone architectures.
- Broad Applicability of Trajectory Signals: The diagnostic results confirmed that historical directional variation serves as an effective and broadly applicable metric for fine-grained guidance allocation.
Potential Applications
The source evaluates TrajCFG across several generative domains without retraining model parameters, including:
- Text-to-video and image-to-video synthesis.
- Text-to-audio-video multimodal generation.
- Text-to-image synthesis.
- Image-to-3D asset generation.
Why This Matters
The findings indicate that historical directional feedback within the denoising trajectory provides a viable, zero-overhead signal for granular guidance allocation. Because TrajCFG reuses standard velocity predictions, it offers a drop-in sampling improvement across multiple backbones and modalities without introducing architectural modifications or retraining overhead.