DeRaDiff — Denoising Time Realignment of Diffusion Models (ICLR 2026)

https://arxiv.org/abs/2601.20198

Official implementation of DeRaDiff, a test-time realignment method for diffusion models. DeRaDiff performs regularized realignment at inference by interpolating between a base diffusion model and a reward-aligned anchor model during the denoising process, controlled by a hyperparameter λ. The framework supports both Stable Diffusion 1.5 and Stable Diffusion XL, and evaluates generated images using CLIP Score, HPS, and PickScore metrics.

Quick Start

1. Environment Setup

conda create -n deradiff python=3.9 -y
conda activate deradiff

2. Install Dependencies

pip install -r requirements.txt
pip install -e .

3. Usage

Place your anchor model checkpoint in the checkpoints/ directory.

Stable Diffusion 1.5:

python scripts/eval_sd15.py \
    --anchor_model_ckpt_path ./checkpoints/anchor_model_ckpt \
    --Lambda 1.0 \
    --Beta 1000 \
    --output_dir ./results/sd15_eval \
    --save_images

Stable Diffusion XL:

python scripts/eval_sdxl.py \
    --anchor_model_ckpt_path ./checkpoints/anchor_model_ckpt \
    --Lambda 1.0 \
    --Beta 1000 \
    --output_dir ./results/sdxl_eval \
    --save_images

Argument	Description
--anchor_model_ckpt_path	Path to the aligned anchor model checkpoint
--Lambda	Realignment strength; approximate regularization is β/λ (default: 1.0)
--Beta	Anchor model regularization strength (default: 1000)
--output_dir	Directory for evaluation results and images
--save_images	Save generated images to disk
--cuda_id	CUDA device ID (default: 0)

Project Structure

DeRaDiff/
├── README.md
├── requirements.txt
├── setup.py
├── checkpoints/             # Anchor model checkpoints
├── data/
│   └── test_prompts.yaml    # 631 evaluation prompts
├── scripts/
│   ├── eval_sd15.py         # Entry point: SD 1.5 evaluation
│   └── eval_sdxl.py         # Entry point: SDXL evaluation
└── src/
    ├── config/
    │   ├── sd15_config.py   # SD 1.5 hyperparameters
    │   └── sdxl_config.py   # SDXL hyperparameters
    ├── eval/
    │   └── evaluator.py     # Multi-metric evaluator (CLIP, HPS, PickScore)
    ├── pipeline/
    │   ├── custom_sd15_pipeline.py   # Custom SD 1.5 denoising pipeline
    │   └── custom_sdxl_pipeline.py   # Custom SDXL denoising pipeline
    └── utils/
        ├── aes_utils.py     # Aesthetic score utilities
        ├── clip_utils.py    # CLIP score computation
        ├── gen_utils.py     # Image generation helper
        ├── hps_utils.py     # HPS score computation
        ├── load_utils.py    # YAML / checkpoint loading
        └── pickscore_utils.py  # PickScore computation