Single Trajectory Distillation for Accelerating Image and Video Style Transfer

Authors: Sijie Xu¹, Runqi Wang^1,2, Wei Zhu¹, Dejia Song¹, Nemo Chen¹, Xu Tang¹, Yao Hu¹
Affiliations: ¹Xiaohongshu, ²ShanghaiTech University

📢 News

• [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:

1. Single Trajectory Distillation Method starting from a specific partial noise state
2. Trajectory Bank for efficient state management
3. 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

pip install -r requirements.txt

Inference Demo (Image-to-Image)

# !pip install opencv-python
import torch
import diffusers
from diffusers import StableDiffusionXLImg2ImgPipeline
from diffusers.schedulers.scheduling_tcd import TCDScheduler
from PIL import Image
 
device = "cuda"
std_lora_path = "weights/std/std_sdxl_i2i_eta0.75.safetensors"
 
pipe = StableDiffusionXLImg2ImgPipeline.from_pretrained("weights/dreamshaper_XL_v21", torch_dtype=torch.float16, variant="fp16").to(device)
 
# load std lora
pipe.scheduler = TCDScheduler.from_config(pipe.scheduler.config, timestep_spacing='leading', steps_offset=1)
pipe.load_lora_weights(std_lora_path, adapter_name="std")
pipe.fuse_lora()
 
# load ipadapter
pipe.load_ip_adapter("ozzygt/sdxl-ip-adapter", "", weight_name="ip-adapter_sdxl_vit-h.safetensors")
pipe.set_ip_adapter_scale(dict(down=0, mid=0, up=dict(block_0=[0, 1, 0], block_1=0))) # only add on 7th block
 
# inputs
prompt = "Stick figure abstract nostalgic style."
n_prompt = "worst face, NSFW, nudity, nipples, (worst quality, low quality:1.4), blurred, low resolution, pixelated, dull colors, overly simplistic, harsh lighting, lack of detail, poorly composed, dark and gloomy atmosphere, (malformed hands:1.4), (poorly drawn hands:1.4), (mutated fingers:1.4), (extra limbs:1.35), (poorly  drawn face:1.4), missing legs, (extra legs:1.4), missing arms, extra arm, ugly, fat, (close shot:1.1), explicit content, sexual content, pornography, adult content, inappropriate, indecent, obscene, vulgar, suggestive, erotic, lewd, provocative, mature content"
src_img = Image.open("doc/imgs/src_img.jpg").resize((960, 1280))
style_img = Image.open("doc/imgs/style_img.png")
 
image = pipe(
    prompt=prompt,
    negative_prompt=n_prompt,
    num_inference_steps=11, # 8 / 0.75 = 11
    guidance_scale=6,
    strength=0.75,
    image=src_img,
    ip_adapter_image=style_img,
).images[0]
 
image.save("std.png")

📦 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.

Image-to-Image Models

η Value	Model Link
0.65	std_sdxl_i2i_eta0.65.safetensors
0.75	std_sdxl_i2i_eta0.75.safetensors
0.85	std_sdxl_i2i_eta0.85.safetensors
0.95	std_sdxl_i2i_eta0.95.safetensors

Video-to-Video Models

η Value	Model Link
0.65	std_sdxl_v2v_eta0.65.safetensors
0.75	std_sdxl_v2v_eta0.75.safetensors
0.85	std_sdxl_v2v_eta0.85.safetensors
0.95	std_sdxl_v2v_eta0.95.safetensors

🛠️ Training

Data Preparation

Download the Open-Sora-Plan-v1 dataset from Hugging Face, which is splitted from Panda70M.

Model Architecture

weights/
├── dinov2_vits14_pretrain.pth
├── ipadapter/
├── motion_adapter_hsxl/
├── open_clip_pytorch_model.bin
└── sdxl_base1.0/

Training Command

bash scripts/std_sdxl_i2i.sh

📚 Citation

@article{xu2024single,
  title={Single Trajectory Distillation for Accelerating Image and Video Style Transfer},
  author={Xu, Sijie and Wang, Runqi and Zhu, Wei and Song, Dejia and Chen, Nemo and Tang, Xu and Hu, Yao},
  journal={arXiv preprint arXiv:2412.18945},
  year={2024}
}

🙏 Acknowledgement

This work builds upon MCM. We thank the open-source community for their valuable contributions.