Fewer Denoising Steps or Cheaper Per-Step Inference: Towards Compute-Optimal Diffusion Model Deployment¶

Conference: ICCV 2025 arXiv: 2508.06160
Code: https://github.com/GATECH-EIC/PostDiff
Area: Diffusion Models / Model Compression Keywords: diffusion model acceleration, mixed-resolution denoising, module caching, training-free compression, compute-optimal deployment

TL;DR¶

This paper proposes PostDiff — a training-free diffusion model acceleration framework that reduces redundancy at two levels: at the input level via a mixed-resolution denoising strategy (low resolution in early steps → high resolution in later steps), and at the module level via a hybrid caching strategy (DeepCache + cross-attention caching). The work systematically addresses the key question of whether reducing the number of denoising steps or reducing the per-step computation cost is more effective — concluding that the latter is superior across most efficiency regimes.

Background & Motivation¶

Background: Diffusion models have achieved remarkable success in image and video generation, but their iterative denoising nature and complex model architectures result in substantial computational cost, limiting deployment on resource-constrained platforms.

Limitations of Prior Work: - Reducing the number of denoising steps (e.g., DDIM, DPM-Solver, consistency models) and reducing per-step computation cost (e.g., token merging/pruning, module caching, quantization) represent two major acceleration paradigms. - However, a systematic study comparing the efficiency–quality trade-offs of these two strategies in the post-training setting has been lacking. - Reducing steps increases the variance of inter-step feature changes, potentially degrading compression compatibility; retaining more steps preserves inter-step redundancy, making compression more applicable.

Key Challenge: In post-training deployment scenarios without fine-tuning, it remains unclear whether fewer denoising steps or cheaper per-step inference is preferable — a question critical to both researchers and practitioners, yet without a definitive answer.

Goal: - Propose a unified framework, PostDiff, that simultaneously reduces redundancy at both the input and module levels. - Systematically compare the two acceleration strategies through controlled experiments. - Identify and explain the "low-resolution enhancement of low-frequency components → improved final quality" phenomenon in mixed-resolution denoising.

Key Insight: Early denoising steps primarily generate low-frequency semantic layouts and do not require high resolution; later steps add high-frequency details and thus benefit from higher resolution. This stage-wise characteristic can be exploited.

Core Idea: Apply low-resolution denoising in early steps (enhancing low-frequency components while saving compute), switch to high resolution in later steps to refine details, and combine with module caching — demonstrating that reducing per-step cost is more effective than reducing the number of steps.

Method¶

Overall Architecture¶

PostDiff consists of two complementary training-free techniques: (1) a mixed-resolution denoising strategy at the input level — early denoising steps operate on low-resolution latents, switching to full resolution at a designated step; and (2) a hybrid module caching strategy at the module level — combining DeepCache (caching the deep skip branches of U-Net) and cross-attention caching (caching conditional guidance information) for reuse across steps.

Key Designs¶

Mixed-Resolution Denoising
- Function: Dynamically switches input resolution during denoising — low resolution in early steps, high resolution in later steps.
- Mechanism: Initializes a low-resolution latent \(x_T^l\) of shape \((\beta w, \beta h)\), where \(0 < \beta < 1\) is the scaling factor. At step \(t = sT\), the method switches to high resolution: the low-resolution prediction \(\hat{x}_0^{l,t}\) is computed via Eq.(2), upsampled via bilinear interpolation as \(\hat{x}_0^{f,t} = \text{Upsample}(\hat{x}_0^{l,t})\), and then mapped to a full-resolution latent via the forward diffusion formula \(x_t^f = \sqrt{\alpha_t}\hat{x}_0^{f,t} + \sqrt{1-\alpha_t}\epsilon\), after which denoising proceeds normally.
- Design Motivation: Early denoising steps predominantly govern low-frequency semantic layout generation, for which low resolution is both sufficient and beneficial — prior literature and empirical results both indicate that low-resolution early steps enhance low-frequency components, ultimately improving final generation quality (a win-win). This is consistent with the empirical success of cascaded diffusion models, where low-frequency information from the low-resolution stage benefits the high-resolution stage.
Visualization of the "Low-Resolution Enhancement" Phenomenon
- Function: Visualizes the per-step evolution of CLIP Score throughout the denoising process.
- Mechanism: Compares full-resolution denoising against mixed-resolution denoising under various settings of \(s\). Observations include: (a) an appropriate number of low-resolution steps can improve the final CLIP Score; (b) although low-resolution steps yield lower CLIP Scores initially, the score recovers more rapidly after switching to high resolution and ultimately surpasses the full-resolution baseline; (c) the balance between low- and high-resolution steps is critical.
- Design Motivation: This visualization not only explains why the method works (low-resolution steps genuinely enhance low-frequency components) but also provides intuition for selecting the hyperparameter \(s\).
Hybrid Module Caching
- Function: Reuses computation at the module level to reduce redundancy.
- Mechanism: Combines two caching strategies: (a) DeepCache — caches the deep skip branches of the U-Net and reuses them for the next \(k=2\) steps; (b) cross-attention caching — disables CFG after step \(m\) and caches the conditional cross-attention, adopting the "Cond" mode as the cache (\(CA_{cache} = CA_t^c\)), which experiments show to be most effective. Key insight: CFG primarily determines layout in early steps and becomes redundant later; moreover, cross-attention can be precomputed during the low-resolution phase for reuse in subsequent steps.
- Design Motivation: (a) High similarity between inter-step feature maps provides the basis for effective caching; (b) cross-attention primarily conveys text-guided layout information, which stabilizes after early steps and can thus be reused; (c) the two caching strategies are complementary — DeepCache reduces spatial computation while cross-attention caching reduces conditional guidance computation.

Loss & Training¶

PostDiff is entirely training-free — no fine-tuning or distillation is required, and it can be applied directly to pretrained diffusion models. Hyperparameters are determined efficiently via a small calibration set whose performance is shown to be highly correlated with that of the full evaluation set.

Core hyperparameters: - \(\beta\): low-resolution scaling factor (1/2 for SD V1.5; 3/4 for SDXL/PixArt) - \(s\): low-to-high resolution switching point (typically 1/2 or 1/5) - \(m\): step at which CFG is disabled (typically 5–15)

Key Experimental Results¶

Main Results: Performance Across Multiple SOTA Diffusion Models¶

Model	Steps	Mix	Cache	FID ↓	CLIP Score ↑	FLOPs (T) ↓	Latency (s) ↓
SD V1.5	20			18.42	30.80	30.420	2.930
SD V1.5	8			20.60	30.41	12.168	1.298
SD V1.5	20	✓		15.69	30.78	19.035	1.945
SD V1.5	20	✓	✓	16.65	30.25	11.129	1.139
SDXL	20			14.10	31.95	119.641	6.521
SDXL	8			18.01	30.92	47.856	2.843
SDXL	20	✓	✓	14.18	31.11	52.682	3.119
PixArt-α	20			29.16	30.41	85.640	7.093
PixArt-α	8			33.09	30.21	34.256	3.031
PixArt-α	20	✓	✓	25.44	30.23	54.718	4.768

Notable finding: on SD V1.5, PostDiff achieves a 63.14% reduction in FLOPs while simultaneously improving FID by 1.8.

Ablation Study: Cross-Attention Caching Strategy Comparison¶

Configuration	FLOPs (T) ↓	FID ↓	CLIP Score ↑
Original	30.420	18.42	30.80
DeepCache (DC)	17.787	17.79	30.75
DC + CA (m=5, Ave)	11.610	18.77	28.40
DC + CA (m=5, Cond)	11.610	18.82	29.20
DC + CA (m=5, CFG)	11.610	103.71	18.22
DC + CA (m=10, Cond)	15.061	21.26	30.11
DC + CA (m=15, Cond)	16.360	21.67	30.37

Comparison with other training-free methods (SD V1.5):

Method	FID ↓	CLIP Score ↑	Latency (s) ↓
Original	18.42	30.80	2.930
DeepCache	17.79	30.75	1.737
TGATE	19.51	29.55	1.992
ToMe	17.43	30.55	2.730
PostDiff	16.65	30.25	1.139

Key Findings¶

Reducing per-step cost > reducing the number of steps: When the goal is to maintain high generation quality (FID < 20), retaining more steps while using PostDiff to reduce per-step cost is superior. Only when pursuing extreme efficiency (> 60% FLOPs reduction) does reducing the number of steps become more favorable.
Mixed-resolution is a win-win: An appropriate number of low-resolution early steps not only saves compute but also improves final FID by enhancing low-frequency components (optimal FID on SD V1.5 improves from 18.42 to 15.69).
CFG caching should not be applied too early: Completely disabling CFG at \(m=5\) (CFG mode) causes quality collapse (FID > 100); the "Cond" mode remains stable across all settings.
Cross-architecture generality: PostDiff is effective across U-Net and Transformer architectures, large and small models, and both LDM and LCM variants.
PostDiff achieves the lowest latency: 1.139s on SD V1.5 versus 1.737s for the next best method, DeepCache (−34%).

Highlights & Insights¶

Systematic answer to the "fewer steps vs. cheaper steps" question: Prior work lacked a fair comparison of these two acceleration paradigms. PostDiff, as a unified framework, enables controlled experiments and yields the actionable conclusion that per-step cost is more important — directly informing deployment decisions.
Low-resolution enhancement of low-frequency components: This is more than an engineering trick; it reflects a deeper insight into the diffusion process — low-resolution denoising in early steps effectively "forces" the model to focus on low-frequency semantic structures, reducing interference from high-frequency noise, consistent with the principles underlying the success of cascaded diffusion models.
Fully training-free: No fine-tuning or distillation is required; PostDiff can be applied plug-and-play to any pretrained diffusion model, making it highly practical.

Limitations & Future Work¶

The paper adopts a simple binary resolution schedule (low → high); more sophisticated progressive resolution schedules may yield further improvements.
Cross-attention caching relies on the CFG mechanism and its applicability to models that do not use CFG (e.g., flow matching) remains to be validated.
The combination with training-aware compression methods such as quantization and pruning has not been explored.
Although efficient, the calibration-set-based hyperparameter selection still incurs some computational overhead.
Validation is limited to image generation; video diffusion models may exhibit different redundancy patterns.

vs. DeepCache: DeepCache caches only the U-Net skip branches; PostDiff additionally incorporates cross-attention caching and mixed-resolution denoising, with all three components jointly pushing the efficiency–quality frontier.
vs. TGATE: TGATE completely disables CFG after step \(m\), which is overly aggressive and degrades quality; PostDiff's cached cross-attention is more conservative, preserving partial conditional information.
vs. ToMe/ToDo: These methods reduce redundancy at the token level (merging/pruning); PostDiff operates at the resolution level — the two approaches are orthogonal and potentially composable.
vs. few-step diffusion models (LCM, consistency models): These require additional training or distillation costs; PostDiff is entirely training-free and can also be combined with LCM (validated experimentally).

Rating¶

Novelty: ⭐⭐⭐⭐ The mixed-resolution denoising strategy is elegant and effective; the research angle of systematically comparing two acceleration paradigms is valuable.
Experimental Thoroughness: ⭐⭐⭐⭐⭐ Four SOTA models × multiple configurations; detailed FID–FLOPs trade-off curves; comparison against 6+ methods.
Writing Quality: ⭐⭐⭐⭐ The research question is well-focused, the experimental design is sound, and the step-wise CLIP Score visualization is compelling.
Value: ⭐⭐⭐⭐ A practical training-free acceleration solution combined with systematic guidance on diffusion model deployment strategies.