ADG: Angle Domain Guidance: Latent Diffusion Requires Rotation Rather Than Extrapolation

This repository contains the official implementation of ADG (Angle Domain Guidance), as presented in our paper accepted at ICML 2025.

📄 Paper: Angle Domain Guidance: Latent Diffusion Requires Rotation Rather Than Extrapolation
📍 Conference: International Conference on Machine Learning (ICML), 2025

---

🔧 Overview

Our method is built upon the diffusers library, offering a plug-and-play interface that requires minimal modification to existing workflows. This repository includes:

• The implementation of the proposed ADG algorithm.
• Example code to run ADG with Stable Diffusion 3.5-large.
• Scripts to reproduce key visualization results from the paper.

---

🚀 Quick Start

1. Set Up Environment

# Create and activate conda environment
conda create -n adg python=3.10 -y
conda activate adg
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118
pip install diffusers transformers accelerate

2. Run ADG

Run the ADG-enhanced sampling pipeline using the following example:

from diffusers import StableDiffusion3Pipeline
from method.ADG_SD3 import ADG_SD3
import torch

# Load the base pipeline
pipeline = StableDiffusion3Pipeline.from_pretrained(
    "stabilityai/stable-diffusion-3.5-large", torch_dtype=torch.float16
)

# Inject ADG into the pipeline
setattr(pipeline, 'ADG', ADG_SD3.__get__(pipeline))

# Generate an image using ADG
prompt = "what you want"
image = pipeline.ADG(
    prompt=prompt,
    num_inference_steps=10,
    guidance_scale=4,
    num_images_per_prompt=1
).images[0]

image.save("output.jpg")

---

📊 Visualization & Reproducibility

To reproduce the key visualizations from the paper and validate experimental results, please refer to vis.ipynb. This notebook demonstrates how ADG enhances consistency and efficiency in diffusion sampling.

---

Note on Earlier Stable Diffusion Models

Thanks to the community feedback for pointing out the low-saturation issue when adapting ADG to SD1.5. We note that ADG is mainly recommended for newer latent diffusion models, such as Stable Diffusion 3 / 3.5. Earlier models such as Stable Diffusion 1.5 may depend more strongly on classifier-free guidance extrapolation to improve saturation and contrast. Since ADG replaces extrapolation with bounded angle-domain rotation, directly applying it to these earlier models may lead to lower-saturation images. We therefore recommend using ADG with more recent diffusion models.

---

📎 Citation

If you find this work helpful, please consider citing:

@inproceedings{jinangle,
  title={Angle Domain Guidance: Latent Diffusion Requires Rotation Rather Than Extrapolation},
  author={Jin, Cheng and Xiao, Zhenyu and Liu, Chutao and Gu, Yuantao},
  booktitle={Forty-second International Conference on Machine Learning}
}