Diffusion transformers (DiTs) now render long, legible text in images, and feature caching is a standard way to accelerate them. Caches are judged by image-level metrics, which do not check whether the rendered text is still correct. We find that caching silently rewrites this text. On Z-Image, TaylorSeer at 2.9x fewer FLOPs lowers word accuracy from 62% to 46%, while ImageReward and CLIP Score barely change. Sampling with fewer steps at the same compute keeps almost all of this text, so the loss comes from how the cache predicts the skipped steps. Decomposing the cache error along classifier-free guidance (CFG), we trace this loss mainly to the guidance term, the difference between the conditional and unconditional predictions, which caches extrapolate worst on text. Correcting the guidance inside the text alone recovers most of the lost text. We therefore propose GuideCache, which recomputes both CFG branches only on the tokens with the largest predicted guidance error, using the guidance terms that the cache already stores. At 2.4x fewer FLOPs, GuideCache restores full-compute prompt fidelity on Z-Image and Qwen-Image, and on Z-Image it renders text as well as or better than ToCa, TeaCache, and FasterCache while staying much closer to the full model's output. GuideCache works with caches of any order, and on top of fewer sampling steps it keeps more text than step reduction at a higher speedup. These results suggest that acceleration methods for text-rendering models should be evaluated on the text they render.
Figures & tables
Figure 2: The text error of the cache comes mainly from the guidance term. (a) Guidance share ρ and extrapolation errors of u and wg on one prompt. (b) Lost prompt fidelity recovered by making one part exact in one region.
Figure 3: Overview of GuideCache. Anchor steps (left) store the shared part ua and the guidance term ga . Cached steps (right) score tokens from the stored ga , recompute both branches on the top- kt tokens, and extrapolate the rest.
Figure 4: Tokens selected by GuideCache on one Z-Image prompt (34 of 50 evaluations), with the predicted guidance error ϵ^t and the selection frequency. Without seeing the OCR boxes (blue), it selects 77% of the tokens inside them and 28% elsewhere.
Qwen-Image
Z-Image
Method
Word ↑
Prompt ↑
PSNR ↑
LPIPS ↓
FLOPs ↑
Lat. ↑
Word ↑
Prompt ↑
PSNR ↑
LPIPS ↓
FLOPs ↑
Lat. ↑
Full (50 steps)
0.605
0.615
∞
0
1.00 ×
1.00 ×
0.621
0.665
∞
0
1.00 ×
1.00 ×
17 anchor evaluations (2.94 × FLOPs)
Step reduction, 17 steps
0.519
0.579
13.5
0.417
2.94 ×
2.87 ×
0.587
0.666
12.9
0.421
2.94 ×
2.91 ×
Step reduction, 22 steps
0.547
0.604
15.2
0.321
2.27 ×
2.24 ×
0.607
0.666
14.0
0.353
2.27 ×
2.26 ×
FasterCache ∗
–
–
–
–
–
–
0.565
0.652
15.4
0.327
1.64 ×
1.70 ×
Table 1: Quantitative comparison on LongText-Bench at 10242 . Methods marked ∗ are our ports to Z-Image, and bold marks the best text score in each block.
Figure 5: Qualitative comparison on Qwen-Image for four randomly drawn prompts, with full images (left) and text crops (right). Sampling with 17 steps and TaylorSeer run at 2.94 × fewer FLOPs, and GuideCache at 2.46 × .
Figure 6: Speed–quality trade-off on Z-Image. Word accuracy (top) and LPIPS (bottom) against measured speedup, with GuideCache at r=5 – 30% and, as blue stars, on a 25-step schedule.
+ GuideCache
Cache
Anchors
FLOPs
Plain
r=20%
r=30%
Full compute
50
1.0 ×
0.645
FORA ( O=0 )
28
1.8 ×
0.641
0.644
0.643
22
2.3 ×
0.630
0.645
0.654
17
2.9 ×
0.622
0.638
0.654
13
3.8 ×
0.475
0.634
0.645
Table 2: GuideCache on caches of any order (Z-Image development subset, prompt fidelity). FLOPs refer to the plain cache.
Choice
Setting
Word ↑
Prompt ↑
FLOPs
Plain cache
0.386
0.601
2.94 ×
Full compute
0.540
0.645
1.00 ×
Selector
Random
0.420
0.623
1.79 ×
ToCa attention score
0.456
0.622
1.79 ×
Guidance share ρa
0.516
0.653
1.80 ×
Shrunk score st
0.532
0.647
1.80 ×
Table 3: Ablation study on the Z-Image development subset with TaylorSeer at 17 anchors and r=30% . Defaults are shaded.
Appendix figures & tables13 assets
Supplementary material from the paper’s appendix.
Appendix
Anchor
Cached
Select
Mem.
Stored g
(s)
(s)
(ms)
(GiB)
(GiB)
Z-Image
Full compute
1.27
–
–
12.6
–
TaylorSeer
1.26
–
–
12.6
–
+ GuideCache
1.27
0.51
0.53
16.4
0.007
Qwen-Image
Appendix
Table A1 : Cost per denoising step at 10242 . Dashes mark parts that a method does not run.
Quantity
Measurement
Setting
Z-Image, 34 of 50 evaluations on the full-compute trajectory (23–24 dB); anchor interval 6 for orders 0–1, 3 for order 2; 48 prompts for pixels, 12 English and 12 Chinese inside the network
Pixel error, text / control
0.96–1.13 × (text / background: 3–4 × )
Latent noise of equal norm
box fidelity of text within 0.02 of the control up to 17% of the latent norm; 0.853 vs. 0.920 at 33%; control SSIM 0.98 to 0.78
Guidance share ∥wg∥/∥v^∥
text 0.50 vs. control 0.32 (1.66 × ); English 1.55 × , Chinese 1.77 ×
Per-token rescaling (Qwen-Image)
TaylorSeer at 17 anchors loses 0.084 word accuracy and 0.018 prompt fidelity with the rescaling, and 0.169 and 0.050 without it (160 prompts, two seeds)
Extrapolation error of g , text / control
reuse 1.83 × , first order 1.96 × , second order 1.30–1.38 ×
Appendix
Table B1: Measurements behind the analysis in Sec. 3.3 .
Exact part
Region
Prompt fid.
Box fid.
Word acc.
Guidance term
text boxes
83%
42%
90%
Shared part
text boxes
11%
2%
2%
Guidance term
control
−2%
−2%
−6%
Both
text boxes
108%
45%
84%
Appendix
Table B2 : Share of the lost text recovered by the causal splice on Z-Image.
Figure B1 : Cache error in text relative to the texture-matched control on Z-Image, as the midpoint and range over cache settings.
Qwen-Image
Z-Image
Method
Word ↑
Prompt ↑
PSNR ↑
LPIPS ↓
FLOPs ↑
Lat. ↑
Word ↑
Prompt ↑
PSNR ↑
LPIPS ↓
FLOPs ↑
Lat. ↑
Stacking on a 25-step schedule (3.1–4.2 × FLOPs)
Step reduction, 15 steps
0.485
0.569
13.0
0.456
3.33 ×
3.20 ×
0.565
0.663
12.5
0.452
3.33 ×
3.30 ×
25 steps + TaylorSeer (12 anchors)
0.512
0.597
15.9
0.295
4.17 ×
3.90 ×
0.461
0.620
13.6
0.380
4.17 ×
4.10 ×
25 steps + TaylorSeer (16 anchors)
0.526
0.604
16.2
0.280
3.12 ×
2.98 ×
0.551
0.657
14.4
0.334
3.12 ×
3.10 ×
25 steps + TaylorSeer (12) + GuideCache ( r=10% )
0.554
0.611
16.1
0.288
3.71 ×
3.36 ×
0.599
0.661
13.9
0.358
3.63 ×
3.50 ×
Appendix
Table C1 : Equal-FLOPs controls, stacking, and CFG truncation , with shaded rows repeated from Tab. 1 .
Qwen-Image
Z-Image
Method
Word ↑
Prompt ↑
PSNR ↑
LPIPS ↓
FLOPs ↑
Lat. ↑
Word ↑
Prompt ↑
PSNR ↑
LPIPS ↓
FLOPs ↑
Lat. ↑
TaylorSeer
0.521
0.597
21.2
0.130
2.94 ×
2.76 ×
0.464
0.625
15.0
0.306
2.94 ×
2.90 ×
+ GuideCache 5%
0.571
0.615
22.3
0.115
2.68 ×
2.22 ×
0.586
0.664
15.7
0.275
2.64 ×
2.50 ×
+ GuideCache 10%
0.583
0.618
23.1
0.104
2.46 ×
2.11 ×
0.606
0.670
16.2
0.256
2.41 ×
2.28 ×
+ GuideCache 20%
0.588
0.617
24.3
0.083
2.14 ×
1.91 ×
0.608
0.670
16.9
0.229
2.07 ×
1.97 ×
+ GuideCache 30%
0.585
0.616
25.3
0.070
1.89 ×
1.71 ×
0.612
0.668
17.8
0.198
1.81 ×
1.70 ×
Appendix
Table C2 : Effect of the recomputed-token ratio on TaylorSeer at 17 anchors, with the 10–30% rows repeated from Tab. 1 .
Qwen-Image
Z-Image
Method
Rewr. ↓
New ↑
Rewr. ↓
New ↑
Step reduction, 17 steps
22.2
14.3
14.2
16.8
Step reduction, 22 steps
15.6
11.8
10.4
14.4
FasterCache ∗
–
–
13.3
8.2
ToCa ∗
–
–
12.1
17.2
TeaCache ∗
–
–
12.4
12.8
Appendix
Table C3 : Rewritten and newly correct strings in the 17-anchor block of Tab. 1 , in percent of the strings that full compute renders correctly or misses.
Method
Word ↑
Prompt ↑
PSNR ↑
LPIPS ↓
FLOPs ↑
Rewr. ↓
TaylorSeer
0.464
0.625
15.0
0.306
2.94 ×
27.7
+ GuideCache ( r=10% )
0.606
0.670
16.2
0.256
2.41 ×
8.2
HiCache
0.491
0.639
15.2
0.298
2.94 ×
22.7
+ GuideCache ( r=10% )
0.596
0.672
16.4
0.255
2.49 ×
9.7
Full (50 steps)
0.621
0.665
∞
0
1.00 ×
–
Appendix
Table C4 : GuideCache on HiCache at 17 anchor evaluations on Z-Image.
Setting
Word ↑
Prompt ↑
LPIPS ↓
FLOPs ↑
λ
0 ( w∥ga∥ only)
0.508
0.629
0.259
2.30 ×
0.25
0.508
0.621
0.252
2.33 ×
0.5
0.521
0.632
0.251
2.41 ×
0.75
0.465
0.631
0.251
2.50 ×
1 ( ϵ^t only)
0.471
0.638
0.259
2.58 ×
W
6, TaylorSeer
0.442
0.616
0.286
2.94 ×
Appendix
Table C5 : Sensitivity to the shrinkage weight λ and the warm-up length W on the Z-Image development subset at r=10% . Defaults are shaded.
Diffusion model caching reuses features, activations, or attention states across denoising steps to reduce inference cost while preserving output quality.
Independent Researcher, Santa Clara, California, United States · Independent Researcher, Mountain View, California, United States · University of Electronic Science and Technology of China