Can a distilled image generator save time by reusing features between its few remaining denoising steps? We study a fixed branch cache that replaces selected attention and MLP evaluations while preserving the student's noise schedule. One full trajectory and one raw-cache trajectory from the same initial noise supply a scalar correction for each branch type and reuse position. At 10242, one internal reuse reduces four-step Qwen-Image-Lightning denoising time from 1.279 to 0.977 seconds. Three reuses yield 1.51× speedup on eight-step Lightning and 1.53× on eight-step Hyper-FLUX. We evaluate image quality on all 200 DrawBench prompts. With the reuse positions held fixed, correction improves mean same-seed image fidelity over raw reuse by 0.66 and 0.87 dB on the two Lightning students. On Hyper-FLUX the PSNR difference is 0.01 dB, while LPIPS improves from 0.1625 to 0.1546. These results measure the effect of branch caching and correction while retaining each student's original sampling grid.
Figures & tables
Student
Denoising (s)
Denoising speedup
Image-ready speedup
ImageReward
CLIP
Lightning-4
1.279 / 0.977
1.309×
1.259×
1.2917 / 1.2980
32.5605 / 32.5591
Lightning-8
2.527 / 1.671
1.513×
1.461×
1.2747 / 1.2922
32.7055 / 32.6572
Hyper-FLUX-8
2.785 / 1.822
1.529×
1.475×
1.0250 / 1.0331
31.5979 / 31.5808
Table 1: Additional speedup over the same uncached student at the same step count. Paired cells show uncached / calibrated means, with the better displayed value in bold. Latency is the mean denoising time measured on NVIDIA H800 GPUs. Speedups use paired H800 measurements. ImageReward and CLIP use all 200 DrawBench prompts.
Figure 1: Branch calibration and reuse. Reproduced from our CalibCache manuscript. The top panel pairs full and raw-cache trajectories from the same initial noise. The lower panel applies the fitted gain to the current branch’s second difference. The diagram illustrates a reuse after an earlier cache step. In the four-step setting the correction runs once, at index 2, after two full evaluations. The eight-step schedule applies it at indices 2, 4 and 6.
Student
Method
Full
Sec. ↓
Speedup ↑
IR ↑
CLIP ↑
PSNR ↑
SSIM ↑
LPIPS ↓
Lightning-4
Native
4
1.284
1.00
1.292
32.56
∞
1.0000
0.0000
TeaCache
3.04
1.030
1.28
1.285
32.66
21.99
0.8290
0.0969
TaylorSeer
3.00
1.045
1.22
1.297
32.51
21.63
0.8418
0.0874
HiCache
3.00
1.034
1.21
1.296
32.52
21.74
0.8424
0.0865
CalibCache
3.00
0.980
1.31
1.298
32.56
22.27
0.8505
0.0783
Lightning-8
Native
8
2.494
1.00
1.275
32.71
∞
1.0000
0.0000
Table 2: Cache comparisons at the original scheduler steps. Quality metrics use DrawBench-200. Latency is the mean denoising time measured on NVIDIA H800 GPUs. Speedup is the ratio of mean times for each method and its own paired native run. Full counts full transformer evaluations per image. PSNR pools RGB squared errors against each same-seed native image before conversion to dB. Bold marks the best reported accelerated values at displayed precision. IR denotes ImageReward.
Student
Raw
Corrected
Difference [95% CI]
Lightning-4
23.30
23.97
+.66[+.59,+.74]
Lightning-8
20.68
21.55
+.87[+.74,+1.01]
Hyper-FLUX-8
22.62
22.63
+.01[−.04,+.07]
Table 3: Paired fidelity relative to each uncached student. Both methods reuse the same branches at the same steps. Each image contributes one dB value, whereas Table 2 pools squared errors. Intervals are paired prompt-bootstrap intervals for corrected minus raw.
Figure 2: “One dog on the street.” DrawBench index 40, seed 314199, from the fixed preview indices. Rows are Lightning-4, Lightning-8 and Hyper-FLUX-8. The broad composition survives in each row. In H8, raw reuse smooths the foreground paving visible in the native image. Images are downsampled for display.
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
Task
Four steps
Eight steps
Object
66.25 / 61.25
65.00 / 56.25
Animal
28.07 / 45.61
24.56 / 21.05
Person
23.29 / 28.77
27.40 / 30.14
Location
48.00 / 52.00
52.00 / 48.00
Background
85.71 / 85.71
85.71 / 100.00
Lighting
100.00 / 100.00
100.00 / 100.00
Appendix
Table 4: Designated-task accuracy (%) on the Mini subset. Columns give native / calibrated Lightning. Each prompt is scored only for its designated task. Macro averages give equal weight to the eight task percentages.
Lightning-4
Lightning-8
Method
Acc.
SSIM
LPIPS
Acc.
SSIM
LPIPS
Native
60.41
1.0000
0.0000
60.83
1.0000
0.0000
TeaCache
62.05
0.8455
0.1093
62.66
0.8178
0.1584
TaylorSeer
62.50
0.8555
0.1030
61.82
0.8221
0.1508
HiCache
63.52
0.8568
0.1011
61.21
0.8167
0.1537
CalibCache
63.50
0.8731
0.0921
61.27
0.8363
0.1379
Appendix
Table 5: DetailMaster Mini subset. Equal-weight mean designated-task accuracy (%) on the identical 50 prompts using the default evaluator. Acc. averages the eight task scores equally. SSIM and LPIPS average over images against the matching native output. Bold compares accelerated methods.
Figure 3: DetailMaster Mini subset image 032, eight-step Lightning. The prompt describes colored lights in a tree with specified positions and attributes. This example was selected to illustrate the detector-threshold case in the text.
Diffusion model caching reuses features, activations, or attention states across denoising steps to reduce inference cost while preserving output quality.
Diffusion models are widely used to generate high-quality images and videos, but their iterative denoising process remains computationally intensive. A growing class of training-free accelerators reduces this cost by reusing cached intermediate features or forecasting future ones. To control draft drift, these methods sometimes compute an exact block feature for verification. Yet the resulting exact feature is typically used only to measure discrepancy or guide a later decision and is then discarded. We find that this previously computed feature can instead be reused for correction. Forwarding it at the verification site resets the local draft residual and reduces downstream feature error. Based on this observation, we introduce FeatFix, a local exact-feature correction method for cached diffusion inference. FeatFix operates at a fixed sparse set of layer--timestep sites. At each selected site, it replaces the complete draft block output with the exact output computed from the same incoming state, avoiding token- or channel-level partial replacement and full-timestep recomputation. Experiments across four image and video backbones show that FeatFix consistently accelerates generation, achieving a speedup of up to 6.70× over Vanilla while maintaining competitive output quality.
Hanshuai Cui, Zhiqing Tang, Zhi Yao +3
School of Artificial Intelligence, Beijing Normal University, Beijing 100875, China · Institute of Artificial Intelligence and Future Networks, Beijing Normal University, Zhuhai 519087, China
Modern diffusion models generate high-quality images and videos, but their iterative denoising process makes inference expensive. Feature caching accelerates sampling by reusing or predicting intermediate activations across neighboring denoising steps, exploiting the redundancy of computations along the reverse trajectory. In this work, we focus on the caching schedule: selecting which denoising steps should be fully recomputed. Existing schedules are either fixed (e.g. uniform) or chosen adaptively from per-step error heuristics; in both cases, the actual compute cost is a side-effect of hand-tuned thresholds rather than a quantity the user can specify. We propose ReCache, which inverts this: given a target budget k, it learns the recomputation schedule that maximizes generation quality, turning compute into a directly controllable input. ReCache trains via policy gradients, sidestepping backpropagation through full diffusion inference, and uses no labelled data. Generations from uncached inference serve as matching targets, paired with a reward for generation quality. ReCache is compatible with any caching mechanism, including feature reuse and feature forecasting; for each mechanism, a single trained policy adapts across computational budgets at inference time. ReCache consistently outperforms scheduling baselines: under a ×5.04 FLOPs reduction on FLUX, it reduces LPIPS by 31% (from 0.456 to 0.316) compared to DiCache; on Wan 2.1 at a ∼×2.6 speedup, it drops LPIPS by 65% (from 0.480 to 0.169) and boosts the VBench score by 7% (5.6 points, from 70.4 to 76.0) over uniform HiCache. Code is available at https://github.com/thecrazymage/ReCache.
Mishan Aliev, Eva Neudachina, Ilya Bykov +4
1HSE University, Russia · 2Yandex Research, Russia
Diffusion and flow matching models achieve impressive visual generation quality but remain computationally expensive due to iterative sampling. Caching-based acceleration provides a practical training-free solution by reusing or predicting intermediate features. However, existing methods largely enforce a forward-only information flow: approximations produced for skipped model evaluations are immediately committed to the sampling trajectory and cannot benefit from feedback generated by later full-computation steps. We propose PoSTCache (Postdictive Smoothing Trajectory Cache), a training-free and plug-and-play framework that reformulates diffusion caching as revisable trajectory inference. PoSTCache first uses an existing caching strategy to construct forward proposals for skipped steps. At a later full-computation step, the discrepancy between the refreshed prediction and its forward proposal yields a correction signal, which PoSTCache propagates backward to refine and reintegrate the preceding predictions. Viewing full-computation steps as sources of corrective feedback also changes how they should be selected. We therefore introduce the Posterior Variance Reduction (PVR) strategy, which estimates cross-timestep dependencies from offline denoising trajectories and selects anchors according to their expected reduction in surrogate trajectory uncertainty rather than local approximation error alone. Experiments on representative image and video generation models show that PoSTCache consistently improves the quality--efficiency trade-off over representative caching methods.
Liwei Lan, Hui Chen, Jungong Han +1
School of Software, Tsinghua University · BNRist, Tsinghua University · Department of Automation, Tsinghua University