Research map / Image generation / Fast diffusion sampling

Accelerating Rectified Flow Models via Trajectory-Aware Caching

NeurIPS 2026 · Image generation › Fast diffusion sampling · approach: Training-free diffusion acceleration · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes TACache, a training-free caching method for rectified flow models that decomposes velocity acceleration into parallel and orthogonal components, schedules skips offline via cumulative variation thresholds, and reconstructs skipped velocities online without extra model evaluations.

Problem

Cache-based acceleration of diffusion and rectified flow samplers skips model evaluations, but coarse approximations accumulate errors over long skip intervals and degrade quality under aggressive acceleration.

Approach

Decompose the discrete velocity acceleration along the current velocity into a parallel magnitude term and an orthogonal directional residual; calibrate population-level magnitude and direction indicators offline to set skip lengths via cumulative-variation thresholds, and at skipped steps reconstruct velocity from these indicators plus the sample's own historical orthogonal direction.

Key results

• On BAGEL, TACache reaches 4.14x speedup with GenEval 87.32% vs 86.31% for the full-step baseline, and PSNR 15.61 vs 15.56 for MagCache.
• On FLUX.1-dev at 3.99x speedup, TACache reports PSNR 22.57 dB, over 4 dB above the next-best accelerated method (18.41 dB for A2S).
• On Wan2.1-1.3B, TACache achieves 2.11x speedup with PSNR 25.19 dB and VBench 72.64 vs 72.59 for the full-step baseline.

Surprising findings

• TACache's GenEval score on BAGEL slightly exceeds the full-step baseline (87.32% vs 86.31%).
• Ablation shows the magnitude indicator (MI) adds marginal gain alone, while the direction indicator (DI) provides the main correction.

Limitations

• Gains on non-reference metrics (CLIP-IQA, Q-Align, GenEval, VBench) sometimes trail competing methods.
• Requires an offline calibration stage on a prompt set to compute indicators and skip schedule.
• Evaluated only on three backbones; longer video generation is left to future work.

Experimental setup

Datasets: WISE InternVid-10M-FLT

Benchmarks: GenEval VBench

Metrics: GenEval accuracy VBench total score PSNR SSIM LPIPS CLIP-IQA Q-Align speedup latency PFLOPs

Baselines: TeaCache TaylorSeer MagCache A2S step truncation (T=13)

Builds on: TeaCache MagCache TaylorSeer EasyCache DiCache A2S

Closest papers by meaning

• MeanCache: From Instantaneous to Average Velocity for Accelerating Flow Matching Inference (ICLR 2026 · similarity 0.92)
• Denoising as Path Planning: Training-Free Acceleration of Diffusion Models with DPCache (CVPR 2026 · similarity 0.90)
• SCoT: Unifying Consistency Models and Rectified Flows via Straight-Consistent Trajectories (NeurIPS 2025 · similarity 0.89)
• MagCache: Fast Video Generation with Magnitude-Aware Cache (NeurIPS 2025 · similarity 0.88)
• Shortcutting Pre-trained Flow Matching Diffusion Models is Almost Free Lunch (NeurIPS 2025 · similarity 0.88)
• Improving the Training of Rectified Flows (NeurIPS 2024 · similarity 0.88)