Research map / Image generation / Fast diffusion sampling

BlockDance: Reuse Structurally Similar Spatio-Temporal Features to Accelerate Diffusion Transformers

CVPR 2025 · Image generation › Fast diffusion sampling · approach: Training-free diffusion acceleration · 3 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Training-free BlockDance caches and reuses structurally similar spatio-temporal block features across adjacent denoising steps in DiTs, plus BlockDance-Ada, an RL-trained decision network choosing per-instance cache/reuse policies, achieving 25-50% speedups.

Problem

Diffusion transformers (DiTs) are slow at inference because of the iterative denoising process, and existing feature-reuse methods do not distinguish features by similarity and scale, causing structural distortions.

Approach

Reuse only the most structurally similar block features (shallow/middle blocks in later denoising steps, after structure stabilizes) via alternating cache and reuse steps, and learn an instance-specific reuse policy with policy gradient using a quality-plus-computation reward.

Key results

• BlockDance (N=2) accelerates DiT-XL/2 on ImageNet by 37.4% with FID 15.70 versus 15.89 for the original model, and reaches 57.5% speedup at N=4.
• On PixArt-α COCO2017, BlockDance (N=2) gives 25.4% latency reduction with FID 30.69 (original 30.41) and SSIM 0.89 vs 0.18–0.60 for baselines.
• On Open-Sora MSR-VTT, BlockDance (N=2) speeds up generation by 34.8% with FVD 550.22 vs 548.72 original, while DeepCache at N=2 degrades FVD to 942.06.

Measured effects reported in the paper

Surprising findings

• Applying feature reuse in early denoising steps (0-40%) introduces structural artifacts, so reuse is only effective after structure stabilizes.
• Reusing deeper blocks loses fine texture detail, so reuse is restricted to shallow and middle blocks.

Limitations

• Limited speedup benefit with very few denoising steps (1 to 4 steps) due to reduced similarity between adjacent steps.
• BlockDance-Ada requires training a decision network per model (PixArt-α trained on a 10k prompt subset).

Experimental setup

Datasets: ImageNet COCO2017 MSR-VTT PixArt-α training data subset

Benchmarks: ImageNet COCO2017 MSR-VTT

Metrics: FID sFID IS CLIP score ImageReward (IR) PickScore SSIM FVD KVD latency

Baselines: ToMe DeepCache TGATE PixArt-LCM

Builds on: DeepCache DDIM DPM-Solver ToMe

Closest papers by meaning

• HarmoniCa: Harmonizing Training and Inference for Better Feature Caching in Diffusion Transformer Acceleration (ICML 2025 · similarity 0.90)
• FasterDiT: Towards Faster Diffusion Transformers Training without Architecture Modification (NeurIPS 2024 · similarity 0.89)
• LazyDiT: Lazy Learning for the Acceleration of Diffusion Transformers (AAAI 2025 · similarity 0.89)
• Rethinking Cross-Layer Information Routing in Diffusion Transformer (NeurIPS 2026 · similarity 0.89)
• SparseDiT: Token Sparsification for Efficient Diffusion Transformer (NeurIPS 2025 · similarity 0.89)
• Accelerating Diffusion Transformers with Token-wise Feature Caching (ICLR 2025 · similarity 0.89)