Reference-to-video diffusion transformers pack the variational autoencoder (VAE) latents of all references and the generated tokens into one full-attention sequence, in which reference tokens make up about half. Because attention cost grows quadratically with sequence length, these tokens take up an even larger share of the attention computation, yet most of them are static, smooth or irrelevant to the output. To accelerate reference-conditioned generation, we propose pruning and caching of reference tokens for video generation (PCRG), a training-free, plug-and-play method that uses two signals the model already computes to assign each reference token to one of three tiers, namely pruned, cached or kept. The first is a low-level redundancy r from VAE latents, since a token whose latent barely changes over time or across its neighbors adds little new content. The second is a high-level importance g from the vision-language-model (VLM) conditioner, where the attention that the prompt and any first or last frames pay to a token tells whether the output needs to read it. Tokens that are not redundant are kept, and redundant tokens are cached if important and pruned otherwise. Pruned tokens leave the sequence before the first step with all rotary position embedding positions unchanged. Cached tokens have their keys and values computed once and reused, and kept tokens are processed as usual. On MiniMax-H3, already distilled to 8 timesteps and quantized to int8, PCRG is 1.57x faster end to end on our V2V-Bench and 1.49x faster on the public VACE-Benchmark, while the quality of VBench drops by only 0.24% and 0.17%, respectively. PCRG shortens the sequence rather than changing how attention is computed or which steps are run, so it also stacks on step caching and sparse attention. On V2V-Bench it reaches 1.86x with TeaCache, 1.82x with Spark at nearly unchanged fidelity and 2.15x with Veda.
Figures & tables
Figure 2: Overview of PCRG. (a) Before step 1, the redundancy r of the VAE latents selects the redundant set R , the most redundant fraction ρ of each block. The importance g from the VLM marks the cells with g>σ as attended. (b) Pruned rows leave the packed sequence, and the remaining tokens keep their RoPE (①). Step 1 computes every remaining row (②). Steps 2– N skip the cached rows and reuse their keys and values from step 1 (③). In the attention maps (④), from step 2 on, a cached token keeps its key column but loses its query row, and a pruned token has neither.
Method
Speed
PSNR ↑
SSIM ↑
LPIPS ↓
Judge ↑
VB-Q
V2V-Bench
Unpruned
1.00 ×
–
–
–
–
84.23
TeaCache [ 10 ]
1.47 ×
16.64
0.569
0.381
+.18
84.44
Spark
1.26 ×
23.46
0.769
0.118
−.04
84.22
Cache-all
1.66 ×
16.25
0.562
0.394
.00
84.10
ToPi [ 23 ]
1.59 ×
17.55
0.595
0.321
−.07
84.15
Table 1: Accelerators on V2V-Bench and VACE-Benchmark. Speed, PSNR, SSIM, and LPIPS is end to end against the ground truth. Judge is the pairwise VLM score against the GT. VB-Q is the mean of six VBench dimensions in percent; changes quoted in the text are computed before rounding. The Bold and UNDERLINE denote the best and second-best results, respectively.
Figure 3: Qualitative comparison on a V2V-Bench request. GT is the unpruned output, and the number in parentheses is the LPIPS to it. +TeaCache and +Spark combine PCRG with TeaCache and Spark.
Selection
Speed
PSNR ↑
SSIM ↑
LPIPS ↓
Prune only, 60% of the cells
by r
1.57 ×
17.89
0.613
0.291
random
1.56 ×
17.54
0.605
0.307
inverted r
1.56 ×
17.28
0.595
0.315
Cache and prune, 70% of the cells
by r (PCRG)
1.65 ×
17.31
0.598
0.317
Table 2: Redundant-set selection on VACE-Benchmark with a reference video: by r , at random, or by inverted r . The Bold and UNDERLINE denote the best and second-best results, respectively.
Figure 4: Redundant-set selection on a layout-control request of VACE-Benchmark The white dashed line is the runner outline from the layout control input, drawn at the same place in every panel. The number in parentheses is the LPIPS to GT.
Pruned cells
PSNR ↑
SSIM ↑
LPIPS ↓
Least attended first
by g
18.18
0.621
0.317
by FL-frame → ref. only
18.04
0.617
0.314
by prompt → ref. only
18.00
0.618
0.324
Most attended first
by g
17.25
0.594
0.367
Table 3: Pruning by importance on the V2V-Bench with an FL-frame. The Bold and UNDERLINE denote the best and second-best results, respectively.
Figure 5: Pruning 60% of the reference cells by importance, no cache. Red boxes means the least attended first dropped. Blue boxes means the most attended first dropped. yellow boxes is GT and PCRG.
Split
Speed
PSNR ↑
SSIM ↑
LPIPS ↓
Cache all redundant
1.27 ×
18.11
0.618
0.281
σ=0.2
1.48 ×
17.63
0.607
0.304
σ=0.4 (PCRG)
1.65 ×
17.31
0.598
0.317
σ=0.6
1.69 ×
17.20
0.593
0.323
σ=0.8
1.70 ×
17.26
0.594
0.321
prune all redundant
1.71 ×
17.19
0.593
0.324
Table 4: Splitting the redundant set of PCRG between cache and prune tiers by σ , on VACE-Benchmark with a reference video.
Configuration
Speed
PSNR ↑
SSIM ↑
LPIPS ↓
DINO ↑
Unpruned
1.00 ×
–
–
–
0.614
Images only
1.06 ×
21.21
0.686
0.190
0.610
Video only
1.43 ×
18.43
0.615
0.290
0.605
PCRG
1.56 ×
18.09
0.607
0.307
0.605
Every redundant image cell
Pruned
1.56 ×
18.07
0.604
0.308
0.607
Table 5: Pruning reference images on the V2V-Bench requests with reference images. Images only and video only apply PCRG to one kind of reference, and the last two rows prune or cache every redundant image cell instead of splitting them by σ . The Bold and UNDERLINE denote the best and second-best results, respectively.
Video diffusion models have recently enabled high-quality video generation with ViT-based architectures, but remain computationally intensive because generation requires attention computation over long spatiotemporal sequences. Token pruning has proven effective for ViTs and VLMs. However, most prior pruning methods are attention-based and operate per frame, failing to ensure the vital temporal coherence across frames in video generation tasks. In practice, naively adopting attention-only pruning causes noticeable degradation due to worsened background consistency, flickering, and reduced image quality. To address this, we propose TAPE, a training-free Temporal Aware Pruning for Efficient diffusion-based video generation. TAPE (i) applies temporal smoothing to align token-importance across adjacent frames and suppress selection jitter; and (ii) performs token reselection in selected layers to align token pruning with layers' diverse semantic focus and avoid error accumulation in specific areas; it also (iii) adopt a timestep-level budget scheduling that prunes aggressively at early noisy steps and relaxes pruning during fidelity-critical refinement. The experimental results show that TAPE delivers significant speedups while preserving high visual fidelity, outperforming prior token reduction approaches.
Sheng Li, Yang Sui, Junhao Ran +3
University of Pittsburgh · Rice University · Rutgers University +1
Few-step autoregressive (AR) video diffusion models generate videos chunk by chunk and have made efficient video generation practical. However, further accelerating these models toward real-time generation remains challenging, as few-step distillation leaves little cross-step redundancy for existing training-free caching methods to exploit. We revisit this regime and identify two token-level redundancy patterns: cross-frame redundancy from static content shared across adjacent frames, and cross-step redundancy from tokens that converge early during denoising. Crucially, we find that early clean estimates along the model's own denoising trajectory provide an effective online signal for locating both patterns, requiring no retraining, auxiliary predictor, or reference video. Based on these findings, we propose PreSTP (Predictive Spatio-Temporal Token Pruning), a training-free framework that constructs a unified token-wise redundancy mask from cross-frame similarity and cross-step convergence. Tokens marked as redundant skip subsequent computation under a consistent last-computed-value restoration rule, and a content-adaptive mask reuse strategy further reduces dense probing cost. On 5-second generation, PreSTP achieves 1.52x-3.20x speedups and a better efficiency-fidelity trade-off than coarse-grained caching baselines such as TeaCache and FlowCache. On 30-second generation, it reaches up to 1.96x speedup while improving VBench-Long and suppressing long-horizon drift. Notably, it continues to reduce inference cost even in the extreme one-step regime, where cross-step redundancy vanishes entirely and coarse step-wise caching is no longer applicable.
Video diffusion transformers generate high-quality videos, but attention over long spatiotemporal sequences makes inference computationally expensive. Sparse attention reduces this cost by computing only a subset of query-key interactions exactly, and recent training-free methods further approximate the remaining interactions. These methods, however, degrade quickly at high sparsity, which limits the achievable speedup. We present ESTRA (Extreme Sparsity Through Refined Attention), a training-free acceleration method that preserves generation quality even at extreme sparsity. ESTRA concentrates exact attention computation on a small set of important key-value tokens and approximates the remaining contributions with cluster representatives. To identify these tokens efficiently, it groups similar queries and uses their representatives to score individual keys. A variance-based correction further refines the attention weights of the cluster representatives, improving the accuracy of the approximation. Experiments on Wan2.1-14B and HunyuanVideo-13B show that ESTRA establishes a new quality-efficiency frontier for training-free sparse attention, achieving 3.26× and 4.71× end-to-end speedups at 5% attention density while surpassing Sol-Attn and SVG-EAR in fidelity to dense generation and on most VBench dimensions. Even at 1% density, ESTRA remains more faithful to dense generation than Sol-Attn at 12.5%, while reaching up to 5.39× speedup.