Towards Stabilized and Efficient Diffusion Transformers through Long-Skip-Connections with Spectral Constraints

• URL: http://arxiv.org/abs/2411.17616v3
• Date: Fri, 28 Mar 2025 16:15:19 GMT
• Title: Towards Stabilized and Efficient Diffusion Transformers through Long-Skip-Connections with Spectral Constraints
• Authors: Guanjie Chen, Xinyu Zhao, Yucheng Zhou, Xiaoye Qu, Tianlong Chen, Yu Cheng,
• Abstract summary: Diffusion Transformers (DiT) have emerged as a powerful architecture for image and video generation, offering superior quality and scalability.<n>DiT's practical application suffers from inherent dynamic feature instability, leading to error amplification during cached inference.<n>We propose Skip-DiT, a novel DiT variant enhanced with Long-Skip-Connections (LSCs) - the key efficiency component in U-Nets.
• Score: 51.83081671798784
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Diffusion Transformers (DiT) have emerged as a powerful architecture for image and video generation, offering superior quality and scalability. However, their practical application suffers from inherent dynamic feature instability, leading to error amplification during cached inference. Through systematic analysis, we identify the absence of long-range feature preservation mechanisms as the root cause of unstable feature propagation and perturbation sensitivity. To this end, we propose Skip-DiT, a novel DiT variant enhanced with Long-Skip-Connections (LSCs) - the key efficiency component in U-Nets. Theoretical spectral norm and visualization analysis demonstrate how LSCs stabilize feature dynamics. Skip-DiT architecture and its stabilized dynamic feature enable an efficient statical caching mechanism that reuses deep features across timesteps while updating shallow components. Extensive experiments across image and video generation tasks demonstrate that Skip-DiT achieves: (1) 4.4 times training acceleration and faster convergence, (2) 1.5-2 times inference acceleration without quality loss and high fidelity to original output, outperforming existing DiT caching methods across various quantitative metrics. Our findings establish long-skip connections as critical architectural components for training stable and efficient diffusion transformers.

Related papers

• Electromyography-Based Gesture Recognition: Hierarchical Feature Extraction for Enhanced Spatial-Temporal Dynamics [0.7083699704958353]
  We propose a lightweight squeeze-excitation deep learning-based multi stream spatial temporal dynamics time-varying feature extraction approach. The proposed model was tested on the Ninapro DB2, DB4, and DB5 datasets, achieving accuracy rates of 96.41%, 92.40%, and 93.34%, respectively.
  arXiv  Detail & Related papers  (2025-04-04T07:11:12Z)
• BlockDance: Reuse Structurally Similar Spatio-Temporal Features to Accelerate Diffusion Transformers [39.08730113749482]
  Diffusion Transformers (DiTs) continue to encounter challenges related to low inference speed. We propose BlockDance, a training-free approach that explores feature similarities at adjacent time steps to accelerate DiTs. We also introduce BlockDance-Ada, a lightweight decision-making network tailored for instance-specific acceleration.
  arXiv  Detail & Related papers  (2025-03-20T08:07:31Z)
• Rethinking Video Tokenization: A Conditioned Diffusion-based Approach [58.164354605550194]
  New tokenizer, Diffusion Conditioned-based Gene Tokenizer, replaces GAN-based decoder with conditional diffusion model. We trained using only a basic MSE diffusion loss for reconstruction, along with KL term and LPIPS perceptual loss from scratch. Even a scaled-down version of CDT (3$times inference speedup) still performs comparably with top baselines.
  arXiv  Detail & Related papers  (2025-03-05T17:59:19Z)
• Q&C: When Quantization Meets Cache in Efficient Image Generation [24.783679431414686]
  We find that the combination of quantization and cache mechanisms for Diffusion Transformers (DiTs) is not straightforward. We propose a hybrid acceleration method by tackling the above challenges. Our method has accelerated DiTs by 12.7x while preserving competitive generation capability.
  arXiv  Detail & Related papers  (2025-03-04T11:19:02Z)
• SmoothCache: A Universal Inference Acceleration Technique for Diffusion Transformers [4.7170474122879575]
  Diffusion Transformers (DiT) have emerged as powerful generative models for various tasks, including image, video, and speech synthesis. We introduce SmoothCache, a model-agnostic inference acceleration technique for DiT architectures. Our experiments demonstrate that SmoothCache achieves 71% 8% to speed up while maintaining or even improving generation quality across diverse modalities.
  arXiv  Detail & Related papers  (2024-11-15T16:24:02Z)
• Adaptive Caching for Faster Video Generation with Diffusion Transformers [52.73348147077075]
  Diffusion Transformers (DiTs) rely on larger models and heavier attention mechanisms, resulting in slower inference speeds. We introduce a training-free method to accelerate video DiTs, termed Adaptive Caching (AdaCache) We also introduce a Motion Regularization (MoReg) scheme to utilize video information within AdaCache, controlling the compute allocation based on motion content.
  arXiv  Detail & Related papers  (2024-11-04T18:59:44Z)
• SparseTem: Boosting the Efficiency of CNN-Based Video Encoders by Exploiting Temporal Continuity [15.872209884833977]
  We propose a memory-efficient scheduling method to eliminate memory overhead and an online adjustment mechanism to minimize accuracy degradation. SparseTem achieves speedup of 1.79x for EfficientDet and 4.72x for CRNN, with minimal accuracy drop and no additional memory overhead.
  arXiv  Detail & Related papers  (2024-10-28T07:13:25Z)
• Faster Diffusion Action Segmentation [9.868244939496678]
  Temporal Action Classification (TAS) is an essential task in video analysis, aiming to segment and classify continuous frames into distinct action segments. Recent advances in diffusion models have demonstrated substantial success in TAS tasks due to their stable training process and high-quality generation capabilities. We propose EffiDiffAct, an efficient and high-performance TAS algorithm.
  arXiv  Detail & Related papers  (2024-08-04T13:23:18Z)
• $Δ$-DiT: A Training-Free Acceleration Method Tailored for Diffusion Transformers [13.433352602762511]
  We propose an overall training-free inference acceleration framework $Delta$-DiT. $Delta$-DiT uses a designed cache mechanism to accelerate the rear DiT blocks in the early sampling stages and the front DiT blocks in the later stages. Experiments on PIXART-$alpha$ and DiT-XL demonstrate that the $Delta$-DiT can achieve a $1.6times$ speedup on the 20-step generation.
  arXiv  Detail & Related papers  (2024-06-03T09:10:44Z)
• PTQ4DiT: Post-training Quantization for Diffusion Transformers [52.902071948957186]
  Post-training Quantization (PTQ) has emerged as a fast and data-efficient solution that can significantly reduce computation and memory footprint. We propose PTQ4DiT, a specifically designed PTQ method for DiTs. We demonstrate that our PTQ4DiT successfully quantizes DiTs to 8-bit precision while preserving comparable generation ability.
  arXiv  Detail & Related papers  (2024-05-25T02:02:08Z)
• TCCT-Net: Two-Stream Network Architecture for Fast and Efficient Engagement Estimation via Behavioral Feature Signals [58.865901821451295]
  We present a novel two-stream feature fusion "Tensor-Convolution and Convolution-Transformer Network" (TCCT-Net) architecture. To better learn the meaningful patterns in the temporal-spatial domain, we design a "CT" stream that integrates a hybrid convolutional-transformer. In parallel, to efficiently extract rich patterns from the temporal-frequency domain, we introduce a "TC" stream that uses Continuous Wavelet Transform (CWT) to represent information in a 2D tensor form.
  arXiv  Detail & Related papers  (2024-04-15T06:01:48Z)
• Faster Diffusion: Rethinking the Role of the Encoder for Diffusion Model Inference [95.42299246592756]
  We study the UNet encoder and empirically analyze the encoder features. We find that encoder features change minimally, whereas the decoder features exhibit substantial variations across different time-steps. We validate our approach on other tasks: text-to-video, personalized generation and reference-guided generation.
  arXiv  Detail & Related papers  (2023-12-15T08:46:43Z)
• Correlated Attention in Transformers for Multivariate Time Series [22.542109523780333]
  We propose a novel correlated attention mechanism, which efficiently captures feature-wise dependencies, and can be seamlessly integrated within the encoder blocks of existing Transformers. In particular, correlated attention operates across feature channels to compute cross-covariance matrices between queries and keys with different lag values, and selectively aggregate representations at the sub-series level. This architecture facilitates automated discovery and representation learning of not only instantaneous but also lagged cross-correlations, while inherently capturing time series auto-correlation.
  arXiv  Detail & Related papers  (2023-11-20T17:35:44Z)
• CAIT: Triple-Win Compression towards High Accuracy, Fast Inference, and Favorable Transferability For ViTs [79.54107547233625]
  Vision Transformers (ViTs) have emerged as state-of-the-art models for various vision tasks. We propose a joint compression method for ViTs that offers both high accuracy and fast inference speed. Our proposed method can achieve state-of-the-art performance across various ViTs.
  arXiv  Detail & Related papers  (2023-09-27T16:12:07Z)
• Latent-Shift: Latent Diffusion with Temporal Shift for Efficient Text-to-Video Generation [115.09597127418452]
  Latent-Shift is an efficient text-to-video generation method based on a pretrained text-to-image generation model. We show that Latent-Shift achieves comparable or better results while being significantly more efficient.
  arXiv  Detail & Related papers  (2023-04-17T17:57:06Z)
• HEAT: Hardware-Efficient Automatic Tensor Decomposition for Transformer Compression [69.36555801766762]
  We propose a hardware-aware tensor decomposition framework, dubbed HEAT, that enables efficient exploration of the exponential space of possible decompositions. We experimentally show that our hardware-aware factorized BERT variants reduce the energy-delay product by 5.7x with less than 1.1% accuracy loss.
  arXiv  Detail & Related papers  (2022-11-30T05:31:45Z)
• Global Vision Transformer Pruning with Hessian-Aware Saliency [93.33895899995224]
  This work challenges the common design philosophy of the Vision Transformer (ViT) model with uniform dimension across all the stacked blocks in a model stage. We derive a novel Hessian-based structural pruning criteria comparable across all layers and structures, with latency-aware regularization for direct latency reduction. Performing iterative pruning on the DeiT-Base model leads to a new architecture family called NViT (Novel ViT), with a novel parameter that utilizes parameters more efficiently.
  arXiv  Detail & Related papers  (2021-10-10T18:04:59Z)
• Feedback-induced instabilities and dynamics in the Jaynes-Cummings model [62.997667081978825]
  We investigate the coherence and steady-state properties of the Jaynes-Cummings model subjected to time-delayed coherent feedback. The introduced feedback qualitatively modifies the dynamical response and steady-state quantum properties of the system.
  arXiv  Detail & Related papers  (2020-06-20T10:07:01Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.