Large Scale Diffusion Distillation via Score-Regularized Continuous-Time Consistency

• URL: http://arxiv.org/abs/2510.08431v1
• Date: Thu, 09 Oct 2025 16:45:30 GMT
• Title: Large Scale Diffusion Distillation via Score-Regularized Continuous-Time Consistency
• Authors: Kaiwen Zheng, Yuji Wang, Qianli Ma, Huayu Chen, Jintao Zhang, Yogesh Balaji, Jianfei Chen, Ming-Yu Liu, Jun Zhu, Qinsheng Zhang,
• Abstract summary: continuous-time consistency model (sCM) is theoretically principled and empirically powerful for accelerating academic-scale diffusion.<n>We first develop a parallelism-compatible FlashAttention-2 JVP kernel, enabling sCM training on models with over 10 billion parameters and high-dimensional video tasks.<n>We propose the score-regularized continuous-time consistency model (rCM), which incorporates score distillation as a long-skip regularizer.
• Score: 60.74505433956616
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: This work represents the first effort to scale up continuous-time consistency distillation to general application-level image and video diffusion models. Although continuous-time consistency model (sCM) is theoretically principled and empirically powerful for accelerating academic-scale diffusion, its applicability to large-scale text-to-image and video tasks remains unclear due to infrastructure challenges in Jacobian-vector product (JVP) computation and the limitations of standard evaluation benchmarks. We first develop a parallelism-compatible FlashAttention-2 JVP kernel, enabling sCM training on models with over 10 billion parameters and high-dimensional video tasks. Our investigation reveals fundamental quality limitations of sCM in fine-detail generation, which we attribute to error accumulation and the "mode-covering" nature of its forward-divergence objective. To remedy this, we propose the score-regularized continuous-time consistency model (rCM), which incorporates score distillation as a long-skip regularizer. This integration complements sCM with the "mode-seeking" reverse divergence, effectively improving visual quality while maintaining high generation diversity. Validated on large-scale models (Cosmos-Predict2, Wan2.1) up to 14B parameters and 5-second videos, rCM matches or surpasses the state-of-the-art distillation method DMD2 on quality metrics while offering notable advantages in diversity, all without GAN tuning or extensive hyperparameter searches. The distilled models generate high-fidelity samples in only $1\sim4$ steps, accelerating diffusion sampling by $15\times\sim50\times$. These results position rCM as a practical and theoretically grounded framework for advancing large-scale diffusion distillation.

Related papers

• Improved Adversarial Diffusion Compression for Real-World Video Super-Resolution [36.32266529540775]
  One-step networks like SeedVR2, DOVE, and DLoRAL are heavy with billions of parameters and multi-second latency.<n>Recent adversarial diffusion compression (ADC) offers a promising path via pruning and distilling these models into a compact AdcSR network.<n>We propose an improved ADC method for Real-VSR that balances spatial details and temporal consistency.
  arXiv  Detail & Related papers  (2026-02-28T04:30:54Z)
• Transition Matching Distillation for Fast Video Generation [63.1049790376783]
  We present Transition Matching Distillation (TMD), a novel framework for distilling video diffusion models into efficient few-step generators.<n>TMD matches the multi-step denoising trajectory of a diffusion model with a few-step probability transition process.<n>TMD provides a flexible and strong trade-off between generation speed and visual quality.
  arXiv  Detail & Related papers  (2026-01-14T21:30:03Z)
• POSE: Phased One-Step Adversarial Equilibrium for Video Diffusion Models [18.761042377485367]
  POSE (Phased One-Step Equilibrium) is a distillation framework that reduces the sampling steps of large-scale video diffusion models.<n>We show that POSE outperforms other acceleration methods on VBench-I2V by average 7.15% in semantic alignment, temporal conference and frame quality.
  arXiv  Detail & Related papers  (2025-08-28T17:20:01Z)
• CHORDS: Diffusion Sampling Accelerator with Multi-core Hierarchical ODE Solvers [72.23291099555459]
  Diffusion-based generative models have become dominant generators of high-fidelity images and videos but remain limited by their computationally expensive inference procedures.<n>This paper explores a general, training-free, and model-agnostic acceleration strategy via multi-core parallelism.<n>ChoRDS significantly accelerates sampling across diverse large-scale image and video diffusion models, yielding up to 2.1x speedup with four cores, improving by 50% over baselines, and 2.9x speedup with eight cores, all without quality degradation.
  arXiv  Detail & Related papers  (2025-07-21T05:48:47Z)
• Dual-Expert Consistency Model for Efficient and High-Quality Video Generation [57.33788820909211]
  We propose a parameter-efficient textbfDual-Expert Consistency Model(DCM), where a semantic expert focuses on learning semantic layout and motion, while a detail expert specializes in fine detail refinement.<n>Our approach achieves state-of-the-art visual quality with significantly reduced sampling steps, demonstrating the effectiveness of expert specialization in video diffusion model distillation.
  arXiv  Detail & Related papers  (2025-06-03T17:55:04Z)
• One-Step Diffusion Model for Image Motion-Deblurring [85.76149042561507]
  We propose a one-step diffusion model for deblurring (OSDD), a novel framework that reduces the denoising process to a single step.<n>To tackle fidelity loss in diffusion models, we introduce an enhanced variational autoencoder (eVAE), which improves structural restoration.<n>Our method achieves strong performance on both full and no-reference metrics.
  arXiv  Detail & Related papers  (2025-03-09T09:39:57Z)
• Latent Consistency Models: Synthesizing High-Resolution Images with Few-Step Inference [60.32804641276217]
  We propose Latent Consistency Models (LCMs), enabling swift inference with minimal steps on any pre-trained LDMs. A high-quality 768 x 768 24-step LCM takes only 32 A100 GPU hours for training. We also introduce Latent Consistency Fine-tuning (LCF), a novel method that is tailored for fine-tuning LCMs on customized image datasets.
  arXiv  Detail & Related papers  (2023-10-06T17:11:58Z)

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.