REVIEW 43 cited by
Delta-DiT: A Training-Free Acceleration Method Tailored for Diffusion Transformers
Not yet reviewed by Pith; the record is open.
This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.
SPECIMEN: schema-true, not a live event
T0 review · schema-true
One-sentence machine reading of the paper's core claim.
pith:XXXXXXXX · record.json · timestamp
Delta-DiT: A Training-Free Acceleration Method Tailored for Diffusion Transformers
read the original abstract
Diffusion models are widely recognized for generating high-quality and diverse images, but their poor real-time performance has led to numerous acceleration works, primarily focusing on UNet-based structures. With the more successful results achieved by diffusion transformers (DiT), there is still a lack of exploration regarding the impact of DiT structure on generation, as well as the absence of an acceleration framework tailored to the DiT architecture. To tackle these challenges, we conduct an investigation into the correlation between DiT blocks and image generation. Our findings reveal that the front blocks of DiT are associated with the outline of the generated images, while the rear blocks are linked to the details. Based on this insight, we propose an overall training-free inference acceleration framework $\Delta$-DiT: using a designed cache mechanism to accelerate the rear DiT blocks in the early sampling stages and the front DiT blocks in the later stages. Specifically, a DiT-specific cache mechanism called $\Delta$-Cache is proposed, which considers the inputs of the previous sampling image and reduces the bias in the inference. Extensive experiments on PIXART-$\alpha$ and DiT-XL demonstrate that the $\Delta$-DiT can achieve a $1.6\times$ speedup on the 20-step generation and even improves performance in most cases. In the scenario of 4-step consistent model generation and the more challenging $1.12\times$ acceleration, our method significantly outperforms existing methods. Our code will be publicly available.
Forward citations
Cited by 43 Pith papers
-
JAGG: Jacobian-Aggregated Group Gradient for Efficient GRPO Training of Diffusion Models
JAGG replaces per-step gradient backpropagation in diffusion GRPO with two endpoint backward passes joined by timestep-weighted interpolation, giving ~2x backward-pass savings at modest quality cost.
-
VDE: Training-Free Accelerating Rectified Flow Model via Velocity Decomposition and Estimation
VDE accelerates rectified flow models like Flux by 3.22x with LPIPS of 0.069 via velocity decomposition into parallel/orthogonal components plus periodic full-pass anchoring.
-
ORBIS: Output-Guided Token Reduction with Distribution-Aware Matching for Video Diffusion Acceleration
ORBIS uses output-guided token reduction and DATM to achieve 2x higher token reduction than AsymRnR, with up to 4.5x speedup and 79.3% energy savings versus A100 GPU for video DiT models.
-
Test-time Sparsity for Extreme Fast Action Diffusion
Test-time sparsity with a parallel pipeline and omnidirectional feature reuse accelerates action diffusion by 5x to 47.5 Hz while cutting FLOPs 92% with no performance loss.
-
1.x-Distill: Breaking the Diversity, Quality, and Efficiency Barrier in Distribution Matching Distillation
1.x-Distill achieves better quality and diversity than prior few-step distillation methods at 1.67 and 1.74 effective NFEs on SD3 models with up to 33x speedup.
-
Not All Frames Deserve Full Computation: Accelerating Autoregressive Video Generation via Selective Computation and Predictive Extrapolation
SCOPE accelerates autoregressive video diffusion up to 4.73x by using a tri-modal cache-predict-recompute scheduler with Taylor extrapolation and selective active-frame computation while preserving output quality.
-
LMEB: Long-horizon Memory Embedding Benchmark
LMEB benchmark shows that embedding models' performance on traditional retrieval does not transfer to long-horizon memory tasks, larger models do not always perform better, and LMEB measures capabilities orthogonal to MTEB.
-
DisCa: Accelerating Video Diffusion Transformers with Distillation-Compatible Learnable Feature Caching
DisCa replaces heuristic feature caching with a lightweight learnable neural predictor compatible with distillation, achieving 11.8× acceleration on video diffusion transformers with preserved generation quality.
-
FeatFix: Reuse What You Verify through Local Exact-Feature Correction for Faster Cached Diffusion Inference
Forwarding the already-computed exact feature at verification sites—instead of discarding it—improves diffusion generation quality at matched compute and reaches up to 6.70× speedup when wrapped around aggressive caches.
-
Importance-Aware OBS Pruning for Diffusion Models
Injecting spatial importance maps (e.g., CFG delta) into the OBS Hessian improves subject preservation in pruned diffusion models at high sparsity, but gains over the baseline are small and without error bars.
-
DiTango: Cost-Effective Parallel Diffusion Generation with Selective Attention State Reuse
Selective reuse of composed attention states across denoising steps lets DiTango skip both remote KV communication and attention compute for low-contribution sequence partitions, cutting multi-GPU diffusion latency by...
-
Kaleido: Algorithm-Hardware Co-Design for Video Diffusion Transformers by Exploiting Latent Space Correlations
A channel-wise reuse algorithm plus a reconfigurable systolic accelerator skips redundant vDiT attention and MLP computation, achieving up to 5.9x speedup and 16x energy savings.
-
Flash-BoN: Instant Drafts for Inference-Time Scaling in Diffusion Models
Under wall-clock budgets, cheap multi-knob drafts plus multi-stage verification outperform guided intermediate search for diffusion T2I inference-time scaling.
-
SAFE-DiT: Semantics-Aware Fast-path Execution for High-Resolution Diffusion Transformers
SAFE-DiT accelerates DiT inference at high resolutions by eliding mask-induced dispatch tax in attention while preserving semantics via selective scheduling, delivering up to 5x speedup and large memory savings with v...
-
RFM-Editing 2: Text-Guided Audio Editing with Rectified Flow Matching and Coarse-to-Fine Diffusion Transformers
Hybrid two-stage diffusion transformer architecture for instruction-guided audio editing via rectified flow that performs joint attention at low resolution then alternates joint and cross-attention at high resolution ...
-
ReCache: Learning Budget-Aware Caching Schedules for Diffusion Models via REINFORCE
ReCache learns recomputation schedules via policy gradients to maximize quality under a target compute budget for any caching mechanism in diffusion models.
-
Light Interaction: Training-Free Inference Acceleration for Interactive Video World Models
Light Interaction accelerates interactive video world models up to 2.59x via adaptive context management, denoising cache acceleration, and 3D block sparse attention without retraining.
-
PARE: Pruning and Adaptive Routing for Efficient Video Generation
PARE applies structure-aware head pruning and timestep/content-conditioned block routing to compress video DiTs, reducing per-step compute while preserving quality on Wan2.1-14B.
-
RT-Lynx: Putting the GEMM Sparsity In a Right Way for Diffusion Models
RT-Lynx shifts DiT sparsity from weights to activations, reports up to 1.55x linear-layer speedup while preserving generation quality across multiple diffusion models.
-
Spectral Progressive Diffusion for Efficient Image and Video Generation
Spectral Progressive Diffusion accelerates image and video generation in pretrained diffusion models by progressively growing resolution along the denoising trajectory using spectral noise expansion and a power spectr...
-
FIS-DiT: Breaking the Few-Step Video Inference Barrier via Training-Free Frame Interleaved Sparsity
FIS-DiT achieves 2.11-2.41x speedup on video DiT models in few-step regimes with negligible quality loss by exploiting frame-wise sparsity and consistency through a training-free interleaved execution strategy.
-
FlashClear: Ultra-Fast Image Content Removal via Efficient Step Distillation and Feature Caching
FlashClear delivers up to 122x faster object removal than prior diffusion models via adversarial step distillation and asymmetric attention caching while preserving visual quality.
-
FlashClear: Ultra-Fast Image Content Removal via Efficient Step Distillation and Feature Caching
FlashClear achieves up to 8.26x speedup over its base diffusion model and 122x over OmniPaint for image object removal via region-aware adversarial distillation and foreground-prioritized caching while claiming to mai...
-
Not All Tokens Need 40 Steps: Heterogeneous Step Allocation in Diffusion Transformers for Efficient Video Generation
HSA assigns variable denoising steps to spatiotemporal tokens in DiTs based on velocity dynamics, with KV-cache sync and cached Euler updates, outperforming prior caching methods on quality-runtime tradeoffs for T2V a...
-
Motion-Aware Caching for Efficient Autoregressive Video Generation
MotionCache accelerates autoregressive video generation up to 6.28x by motion-weighted cache reuse based on inter-frame differences, with negligible quality loss on SkyReels-V2 and MAGI-1.
-
CoCoDiff: Optimizing Collective Communications for Distributed Diffusion Transformer Inference Under Ulysses Sequence Parallelism
CoCoDiff achieves 3.6x average and 8.4x peak speedup for distributed DiT inference on up to 96 GPU tiles via tile-aware all-to-all, V-first scheduling, and selective V communication.
-
DiffSparse: Accelerating Diffusion Transformers with Learned Token Sparsity
A learnable cost predictor plus dynamic programming allocates layer-wise token sparsity for diffusion transformers, removing forced full steps and cutting ~54% compute on PixArt-α without quality loss.
-
LMEB: Long-horizon Memory Embedding Benchmark
LMEB is a 22-dataset, 193-task zero-shot benchmark showing that long-horizon memory retrieval is hard, not solved by scale, and largely orthogonal to MTEB passage-retrieval skill.
-
LMEB: Long-horizon Memory Embedding Benchmark
LMEB is a new benchmark of 193 retrieval tasks spanning episodic, dialogue, semantic, and procedural memory on which top embedding models score about 61 NDCG@10, largely uncorrelated with MTEB.
-
Block-wise Adaptive Caching for Accelerating Diffusion Policy
BAC accelerates transformer-based Diffusion Policy up to 3x by block-level adaptive feature caching using an Adaptive Caching Scheduler and Bubbling Union Algorithm to control error propagation.
-
PipeFusion: Patch-level Pipeline Parallelism for Diffusion Transformers Inference
PipeFusion applies patch partitioning and pipeline parallelism with one-step stale feature reuse to reduce communication overhead in DiT inference, reporting SOTA results on 8x L40 GPUs for Pixart, SD3, and Flux.1.
-
OmniCache: Multidimensional Hierarchical Feature Caching For Diffusion Models
Training-free hierarchical feature caching across token, frame, block, and layer axes cuts diffusion inference latency up to 35% while preserving quality better than averaging-based token merging.
-
Vera: Identity-Faithful Human Subject-to-Video Generation
Vera improves identity consistency in human subject-to-video generation using cross-clip identity-aligned data, face-weighted masked loss, and layer-aware reference attention.
-
ACID: Adaptive Caching for vIDeo generation
ACID dynamically switches caching thresholds based on drift-signal rate of change, recovering near-conservative quality at substantially higher video-diffusion inference speeds.
-
ACID: Adaptive Caching for vIDeo generation
Adaptive threshold switching on the drift signal's rate of change expands the quality-vs-speed Pareto frontier of TeaCache, EasyCache, and DiCache across HunyuanVideo, Wan 2.1, and CogVideoX.
-
OTCache: Optimal Transport for Geometry-Aware Caching in Diffusion Models
OTCache uses optimal transport to interpolate caching schedules between a graph-based reference and an Optuna-optimized anchor, delivering 3.66x-4.7x speedups on FLUX.1, Qwen-Image and HunyuanVideo with improved fidelity.
-
ResilPhase: Plug-and-Play Phase Mapping and Noise-Resilient Macro-Trajectory Extrapolation for Diffusion Acceleration
ResilPhase accelerates DiT inference via stable ODE macro-trajectory extrapolation with derivative-free barycentric Lagrange extrapolator and bounded phase mapping, claiming SOTA fidelity at high acceleration ratios o...
-
RFM-Editing 2: Text-Guided Audio Editing with Rectified Flow Matching and Coarse-to-Fine Diffusion Transformers
Hybrid two-stage diffusion transformer architecture for instruction-guided audio editing uses coarse joint attention at low resolution and refined alternating blocks at high resolution to improve performance and effic...
-
RFM-Editing 2: Text-Guided Audio Editing with Rectified Flow Matching and Coarse-to-Fine Diffusion Transformers
A coarse-to-fine hybrid MMDiT/DiT audio editor trained with rectified flow matching improves fidelity and cuts edit time versus prior instruction-guided baselines on synthetic overlapping-event tasks.
-
Spectral Progressive Diffusion for Efficient Image and Video Generation
Spectral Progressive Diffusion progressively grows resolution during denoising of pretrained diffusion models via spectral noise expansion and a power-spectrum-derived schedule, enabling training-free speedups and a f...
-
Motion-Aware Caching for Efficient Autoregressive Video Generation
MotionCache speeds up autoregressive video generation by 6.28x on SkyReels-V2 and 1.64x on MAGI-1 via motion-weighted cache reuse based on inter-frame differences, with negligible quality loss on VBench.
-
Training-free, Perceptually Consistent Low-Resolution Previews with High-Resolution Image for Efficient Workflows of Diffusion Models
A commutator-zero condition enables training-free generation of perceptually consistent low-resolution previews for high-resolution diffusion model outputs, achieving up to 33% computation reduction.
-
AdaCorrection: Adaptive Offset Cache Correction for Accurate Diffusion Transformers
AdaCorrection adaptively corrects offset caches in DiT inference via on-the-fly spatio-temporal validity checks to maintain near-original FID with moderate acceleration.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.