Golden Noise for Diffusion Models: A Learning Framework

TODO List

• Uploaded inference code and corresponding model weights.
• Uploaded data collection, model training, and evaluation code.
• Upload the latest version of the paper.
• Check for hidden bugs in the codebase.

Preliminaries

This section will guide you through setting up the environment required to run and develop with this project.

1. Python Environment Setup

We recommend using Anaconda or Miniconda to manage your Python environments, but you can also use venv or other tools.

Using Conda (Recommended)

# Create a new environment with Python 3.8+
conda create -n golden-noise python=3.8 -y
conda activate golden-noise

Using venv (Standard Library)

# Create a new virtual environment
python -m venv golden-noise-env
# Activate the environment (Windows)
golden-noise-env\Scripts\activate
# Activate the environment (Linux/MacOS)
source golden-noise-env/bin/activate

2. Install PyTorch

Please follow the official PyTorch installation guide to install the correct version for your CUDA driver. For example:

# Example: install PyTorch with CUDA 11.8
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

3. Install Other Dependencies

Install the required Python packages:

pip install diffusers PIL numpy timm argparse einops

Note: If you encounter issues with package versions, please refer to the requirements in this README or open an issue.

---

Dataset

The data/ folder provides several resources for both training and evaluation:

• drawbench.csv, HPD_prompt.csv, pickscore.csv: Three test sets containing prompts and evaluation data for benchmarking model performance.
• pickscore_train_prompts.json: A prompt dataset used for collecting training data, containing prompts for generating training pairs.
• train_60000.json: An example training dataset, containing prompts and seeds, demonstrating the format and structure for model training.

Data Preparation for Training

To prepare training data, use the script training/dataset_collection.py. This script reads prompts and seeds from the selected prompt dataset (e.g., train_60000.json), generates corresponding source noise and target noise for each prompt, and saves the results into .npz files for efficient training.

---

Training Usage

Notice: Some folders cannot be directly downloaded from GitHub due to repository file size and policy restrictions. We provide external download links for these resources. Please download them manually and place them in the training directory as instructed. Download Link: google drive

Training scripts are located in the training/ directory.

Main Scripts

• main.py: Entry point for model training
• dataset_collection.py: Collects and processes training data by generating source and target noise pairs from prompt datasets. Run this script before training to prepare your .npz training files.

Example: Prepare Training Data

python training/dataset_collection.py --prompt_dataset data/train_60000.json --output_dir data/train_npz/

This will generate .npz files containing the source and target noise for each prompt in the specified dataset.

Example: Start Training (SDXL NPNet)

python training/main.py --pipeline=SDXL --model=svd_unet+unet --train=True --pick=True --all-file=True --discard=True --test=True --postfix=_test --evaluate=False --discard=True

main.py Command-Line Arguments Explanation

• --ddp: Whether to use Distributed Data Parallel (DDP) training. Default: False.
• --pipeline: Model pipeline to use. Options: 'SDXL', 'SD2.1', 'DS-turbo', 'DiT'. Default: 'SDXL'.
• --model: Model architecture. Options: 'unet', 'vit', 'svd_unet', 'svd_unet+unet', 'e_unet', 'svd_unet+unet+dit'. Default: 'svd_unet+unet'.
• --benchmark-type: Benchmark type. Options: 'pick', 'draw'. Default: 'pick'.
• --train: Whether to run training. Default: False.
• --test: Whether to run testing/inference. Default: False.
• --postfix: Postfix for output files and checkpoints. Default: '_hps_sdxl_step_10_random_noise'.
• --acculumate-steps: Number of gradient accumulation steps. Default: 64.
• --pick: Whether to use PickScore for filtering or evaluation. Default: False.
• --do-classifier-free-guidance: Whether to use classifier-free guidance during generation. Default: True.
• --inference-step: Number of inference steps for the diffusion process. Default: 10.
• --size: Image size (height and width). Default: 1024.
• --RatioT: Ratio parameter for training (custom use). Default: 1.0.
• --guidance-scale: Guidance scale for classifier-free guidance. Default: 5.5.
• --guidance-rescale: Rescale factor for guidance. Default: 0.0.
• --all-file: Whether to use all files in the dataset. Default: False.
• --epochs: Number of training epochs. Default: 30.
• --batch-size: Batch size for training. Default: 64.
• --num-workers: Number of worker processes for data loading. Default: 16.
• --metric-version: Metric for evaluation. Options: 'PickScore', 'HPS v2', 'AES', 'ImageReward'. Default: 'PickScore'.
• --prompt-path: Path to the prompt JSON file for training. Default: './sdxl_step_10_training_seed.json'.
• --data-dir: Directory containing the training data (noise pairs). Default: './datasets/noise_pairs_SDXL_10_pick_total/'.
• --pretrained-path: Path to pretrained model weights. Default: './training/checkpoints'.
• --save-ckpt-path: Path to save model checkpoints. Default: './training/checkpoints/SDXL-10/svd_unet+unet'.
• --discard: Whether to discard bad samples during training. Default: False.

Important Setup (Metrics and Reward Models)

1. Download the metric and reward_model folders from Google Drive and place them under the training/ directory.

• Download link: google drive
• Final layout example:

training/
  metric/
  reward_model/
  utils/
  solver/
  ...

Data Collection Details (dataset_collection.py)

• The script will create a JSONL file that records each prompt and its corresponding random_seed. This JSONL will be used later by main.py during training as --prompt-path.
• Saved data (.npz) will contain the paired latents/noise as defined in the code. You can edit cal_score to collect more metrics. The storage structure follows the logic in the code.

Multi-GPU data collection and index management:

• If different GPUs generate different index ranges, adjust the sample index to avoid collisions.
• Example:
  • GPU 0 generates indices [0, 10000): do not change idx.
  • GPU 1 generates indices [10000, 20000): set idx = idx + 10000 before saving.
• After collection, merge all JSONL files into a single file for training. The .npz files do not require merging or further processing.

Model Training Guide (main.py)

• Recommended baseline: --model=svd_unet+unet.
• Use --pick True as the typical setting. To filter bad samples, add --discard True. The filtering logic is implemented in training/utils/utils.py via load_pick_discard_prompt; modify it if needed.
• If you have multiple .npz folders, set --all-file True; otherwise keep it False.
• Set --prompt-path to the JSONL created by dataset_collection.py (if you produced multiple JSONLs, merge them into one file first).

Example command:

python training/main.py \
  --pipeline SDXL \
  --model svd_unet+unet \
  --pick True \
  --discard True \
  --all-file True \
  --prompt-path ./training/SDXL_step_10_training.json \
  --data-dir ./training/datasets/noise_pairs_SDXL_10_pick_total/ \
  --epochs 30 \
  --batch-size 64

Example: Evaluate the model performance

We provide evaluation code in training/metric/cal_metric.py for quantitative assessment of generated images.

Usage Notes:

• You need to prepare the corresponding prompt file and the generated images for evaluation.
• The image folder should follow this structure:

images/
  origin/
    0.png
    1.png
    ...
  optim/
    0.png
    1.png
    ...

• origin/ contains images generated by the baseline model, and optim/ contains images generated by the optimized (e.g., NPNet) model. The file names should correspond to the prompt order.
• Before running the evaluation, make sure you have downloaded the required weights for the reward model you want to use (e.g., PickScore, HPSv2, ImageReward, etc.).

Example command:

python training/metric/cal_metric.py --prompt_file data/your_prompts_file --image_folder images/ --metric PickScore

Replace --metric with your desired evaluation metric and adjust paths as needed.

NPNet Pipeline Inference Usage Guide😄

Overview

This guide provides instructions on how to use the NPNet, a noise prompt network aims to transform the random Gaussian noise into golden noise, by adding a small desirable perturbation derived from the text prompt to boost the overall quality and semantic faithfulness of the synthesized images.

Here we provide the inference code which supports different models like Stable Diffusion XL, DreamShaper-xl-v2-turbo, and Hunyuan-DiT..

Besides, you can apply the checkpoint of NPNet on SDXL to the models like SDXL-Lightning, LCM, and PCM. The visualizations of these three models are shown below:

We directly use the checkpoint from SDXL to SDXL-Lightning, LCM, and PCM, and evaluate them on Geneval dataset:

Model	Method	PickScore↑	HPSv2↑	AES↑	ImageReward↑	CLIPScore↑
SDXL-Lightning(4-step)	standard	22.85	29.12	5.65	59.02	0.8093
SDXL-Lightning(4-step)	ours	23.03	29.71	5.71	72.67	0.8150
LCM(4-step)	standard	22.30	26.52	5.49	33.21	0.8050
LCM(4-step)	ours	22.38	26.83	5.55	37.08	0.8123
PCM(8-step)	standard	22.05	26.98	5.52	23.28	0.8031
PCM(8-step)	ours	22.22	27.59	5.56	35.01	0.8175

The results demonstrate the effectiveness of our NPNet on few-steps image generation.

Usage👀️

To use the NPNet pipeline, you need to run the npnet_pipeline.py script with appropriate command-line arguments. Below are the available options:

Command-Line Arguments

• --pipeline: Select the model pipeline (SDXL, DreamShaper, DiT). Default is SDXL.
• --prompt: The textual prompt based on which the image will be generated. Default is "A banana on the left of an apple."
• --inference-step: Number of inference steps for the diffusion process. Default is 50.
• --cfg: Classifier-free guidance scale. Default is 5.5.
• --pretrained-path: Path to the pretrained model weights. Default is a specified path in the script.
• --size: The size (height and width) of the generated image. Default is 1024.

Running the Script

Run the script from the command line by navigating to the directory containing npnet_pipeline.py and executing:

python npnet_pipeline.py --pipeline SDXL --prompt "A banana on the left of an apple." --size 1024

This command will generate an image based on the prompt "A banana on the left of an apple." using the Stable Diffusion XL model with an image size of 1024x1024 pixels.

Output🎉️

The script will save two images:

• A standard image generated by the diffusion model.
• A golden image generated by the diffusion model with the NPNet.

Both images will be saved in the current directory with names based on the model and prompt.

Pre-trained Weights Download❤️

We provide the pre-trained NPNet weights of Stable Diffusion XL, DreamShaper-xl-v2-turbo, and Hunyuan-DiT with google drive

Citation:

If you find our code useful for your research, please cite our paper.

@inproceedings{zhou2025golden,
      title={Golden Noise for Diffusion Models: A Learning Framework}, 
      author={Zikai Zhou and Shitong Shao and Lichen Bai and Shufei Zhang and Zhiqiang Xu and Bo Han and Zeke Xie},
      booktitle={International Conference on Computer Vision},
      year={2025},
}

🙏 Acknowledgements

We thank the community and contributors for their invaluable support in developing NPNet. We thank @DataCTE for constructing the ComfyUI of NPNet inference code ComfyUI. We thank @asagi4 for constructing the ComfyUI of NPNet inference code ComfyUI.

---