FireRedTeam 开源 Single-Trajectory Distillation：8-4 步高质量风格化蒸馏
GitHub - FireRedTeam/Single-Trajectory-Distillation · GitHub
小红书与上海科技大学团队开源 Single Trajectory Distillation（STD），通过单轨迹蒸馏、Trajectory Bank 和基于 DINO-v2 的非对称对抗损失，实现 8-4 步高质量图像与视频风格化，并可直接接入现有 SDXL 流程。实验显示其在风格相似度和美学指标上超过 LCM、TCD、PCM 等加速模型，预训练模型已发布于 Hugging Face。
正文
当前显示原文，中文全文译文尚未就绪。
FireRedTeam; Single-Trajectory-Distillation
Repository navigation
NotificationsYou must be signed in to change notification settings; Fork0 (0); Star27 (27)You must be signed in to star a repository
Repository files navigation
README; MIT license
Affiliations: 1Xiaohongshu, 2ShanghaiTech University
[2025.02.11] 🔥 Source code released; [2025.02.10] 🔥 Pretrained models available on Hugging Face; [2024.12.31] 🔥 Paper published on arXiv
📖 Abstract
Diffusion-based stylization methods typically denoise from specific partial noise states for image/video tasks. This multi-step process faces computational challenges that hinder real-world applications. While consistency models through trajectory distillation offer acceleration potential, existing approaches only enforce initial-step alignment between student and imperfect teacher models. We propose Single Trajectory Distillation (STD) with three key innovations:
Single Trajectory Distillation Method starting from a specific partial noise state; Trajectory Bank for efficient state management; Asymmetric Adversarial Loss based on DINO-v2 for quality enhancement
Extensive experiments demonstrate our method surpasses existing acceleration models (LCM, TCD, PCM, etc.) in style similarity and aesthetic metrics.
🎯 Key Features
8-4 step high-quality stylization; Unified framework for image & video processin; Plug-and-play integration with existing SDXL pipelines
🖼️ Visual Results
Method Overview
Architecture diagram showing (Left) Trajectory Bank management, (Center) Single-trajectory distillation framework, (Right) Asymmetric adversarial loss component
Qualitative Comparison
Visual comparison with LCM, TCD, PCM, and other baselines at NFE=8 (CFG=6)
Metric Analysis
Performance under different CFG values (2-8). Our method (red line) achieves optimal style-content balance.
🚀 Quick Start
Installation
Inference Demo (Image-to-Image)
📦 Model Zoo
We provide pretrained models for both image-to-image and video-to-video tasks with different η values. All models are hosted on Hugging Face.
Download the Open-Sora-Plan-v1 dataset from Hugging Face, which is splitted from Panda70M.
Model Architecture
Training Command
This work builds upon MCM. We thank the open-source community for their valuable contributions.