Diffusion Transformers generate high-quality images and videos, but every sampling step requires a full network evaluation. Training-free caching reduces this cost by reusing or forecasting network outputs across steps, but common approaches use per-model calibrated skip criteria or maintain per-layer prediction histories. We present D3-Cache, which applies Dynamic Mode Decomposition (DMD) to Diffusion sampling: it fits DMD to a short history of exact velocities and forecasts only the velocity that the scheduler consumes. A horizon-aware gate decides when to skip: whenever the network is evaluated, its output is compared with the forecast for that step, and the larger the error, the sooner the network is evaluated again, up to a cap on consecutive skipped steps. D3-Cache requires no training or per-model offline calibration curve and adds negligible memory. It shares one history length and risk tolerance across the four evaluated models. On four state-of-the-art image and video models, D3-Cache achieves speedups of 2.6-3.8x with high fidelity to the base model output.
Figures & tables
Figure 1: Velocity and patch-correlation diagnostic on FLUX. (a) Absolute difference between the patch-correlation matrices of forecast ( Cmethod ) and exact ( Cbase ) velocities. (b,c) Per-prompt correlation mismatch and velocity error on eight prompts: DMD lowers both relative to first- and second-order TaylorSeer (red: mean reduction relative to O=1 / O=2 ). Controlled comparisons with shared histories are given in the supplement.
Figure 2: Overview of D 3 -Cache. (a) After warmup, each sampling step either runs a full evaluation or reuses a DMD forecast of the velocity. The estimated risk grows with the horizon since the last full evaluation; exceeding τ or reaching the skip cap S triggers a refresh. (b) At a refresh, the exact velocity first scores the forecast made from the preceding history \scriptsize1⃝, which updates the risk slope for later steps, and is then pushed into the history used for the next DMD fit \scriptsize2⃝. For subsequent steps, the gate selects the forecast when Rk(t)≤τ and s<S ; the DiT blocks and output head are then skipped entirely. Otherwise they are evaluated and the exact velocity is selected.
Method
Efficiency
Fidelity
Perceptual
Latency (s) ↓
Speedup ↑
VRAM (GiB) ↓
SSIM ↑
PSNR ↑
LPIPS ↓
rFID ↓
IR ↑
CLIP ↑
Vanilla (50 steps)
63.15
1.000 ×
59.89
–
–
–
–
1.271
28.84
Vanilla (25 steps)
31.75
1.989 ×
59.89
0.6201
14.22
0.3991
67.02
1.247
28.52
TeaCache ( δ=0.36 )
26.96
2.342 ×
59.93
0.7360
18.39
0.2319
43.87
1.231
28.74
MagCache ( δ=0.08,K=4 )
26.95
2.343 ×
59.93
0.7723
19.75
0.1889
36.39
1.241
28.75
TaylorSeer ( N=3,O=2 )
26.52
2.381 ×
72.19
0.6510
15.80
0.3477
60.43
1.198
28.78
Table 1: Quantitative comparison in image generation on HiDream-I1-full (200 DrawBench prompts, 50 steps, 1024 × 1024). Fidelity is measured against the 50-step output; IR: ImageReward. Bold marks the best cache-based method.
Method
Efficiency
Fidelity
Perceptual
Latency (s) ↓
Speedup ↑
VRAM (GiB) ↓
SSIM ↑
PSNR ↑
LPIPS ↓
rFID ↓
IR ↑
CLIP ↑
Vanilla (50 steps)
31.12
1.000 ×
33.81
–
–
–
–
1.077
26.25
Vanilla (25 steps)
15.72
1.980 ×
33.81
0.7581
18.37
0.2831
63.11
1.043
26.34
TeaCache ( δ=0.95 )
8.63
3.604 ×
33.86
0.7484
19.21
0.2934
66.86
1.023
26.15
MagCache ( δ=0.16,K=6 )
8.34
3.733 ×
33.84
0.8019
21.29
0.2190
50.71
1.028
26.13
TaylorSeer ( N=6,O=2 )
8.29
3.753 ×
39.82
0.6977
17.77
0.3317
66.18
1.053
26.24
Table 2: Quantitative comparison in image generation on FLUX.1-dev (100 COCO captions, 50 steps, 1024 × 1024). † : run in its native pipeline; latency normalized to the common 50-step baseline.
Figure 3: Qualitative comparison with previous acceleration methods on HiDream-I1 (DrawBench prompts). Compared methods change the background, the window contents or the pear texture, whereas D 3 -Cache stays closest to the 50-step output. Examples are taken from the prompts with the largest LPIPS advantage of D 3 -Cache; speedups are those of Table 1 .
Method
Efficiency
Fidelity
Quality
Latency (s) ↓
Speedup ↑
VRAM (GiB) ↓
SSIM ↑
PSNR ↑
LPIPS ↓
VBench ↑
Vanilla (50 steps)
275.56
1.000 ×
60.17
–
–
–
82.64
Vanilla (25 steps)
141.56
1.947 ×
60.17
0.7946
22.51
0.2050
81.60
TeaCache ( δ=0.15 )
127.69
2.158 ×
60.52
0.8232
24.54
0.1602
81.64
MagCache ( δ=0.30,K=6 )
104.74
2.631 ×
60.28
0.9007
29.41
0.0871
81.84
TaylorSeer ( N=6,O=1 )
101.85
2.706 ×
70.12
0.8184
26.39
0.1091
82.52
Table 3: Quantitative comparison in video generation on HunyuanVideo (128 VBench prompts, 50 steps, 640 × 480, 65 frames). Second-order TaylorSeer and HiCache run out of GPU memory on this model, so their first-order settings are reported.
Method
Efficiency
Fidelity
Quality
Latency (s) ↓
Speedup ↑
VRAM (GiB) ↓
SSIM ↑
PSNR ↑
LPIPS ↓
VBench ↑
Vanilla (50 steps)
178.38
1.000 ×
24.53
–
–
–
82.65
Vanilla (25 steps)
92.15
1.936 ×
24.53
0.6101
14.31
0.4151
83.01
TeaCache ( δ=0.25 )
68.58
2.601 ×
24.83
0.8574
23.98
0.1178
82.25
MagCache ( δ=0.24,K=6 )
68.55
2.602 ×
24.83
0.8601
24.03
0.1153
82.39
TaylorSeer ( N=7,O=2 )
67.84
2.629 ×
65.50
0.7932
21.15
0.1549
81.40
Table 4: Quantitative comparison in video generation on Wan2.1-T2V-1.3B (128 VBench prompts, 50 steps, 832 × 480, 65 frames).
Figure 4: Qualitative comparison with previous acceleration methods on HunyuanVideo. One frame per clip, with the highlighted region magnified. Compared methods blur or redraw the jogger’s legs, the cat’s face and the sausage, whereas D 3 -Cache stays closest to the 50-step output. TaylorSeer and HiCache use their first-order settings, as the second-order ones exceed the GPU memory. Examples are taken from the prompts with the largest LPIPS advantage of D 3 -Cache.