Training-Free Safe Text Embedding Guidance for Text-to-Image Diffusion Models (STG) (NeurIPS 2025)

| paper | openreview | code | poster, slide, video |

---

This repository contains the official PyTorch implementation of Training-Free Safe Text Embedding Guidance (STG) in NeurIPS 2025.

Byeonghu Na, Mina Kang, Jiseok Kwak, Minsang Park, Jiwoo Shin, SeJoon Jun, Gayoung Lee, Jin-Hwa Kim, and Il-Chul Moon

KAIST, NAVER AI Lab, SNU AIIS, summary.ai

---

Safe Text embedding Guidance (STG) is a training-free text embedding guidance method designed to improve safety in text-to-image diffusion models.

---

Setup

We utilized CUDA 11.4 on a single NVIDIA A100 GPU.

pip install -r requirements.txt

Sampling

Without STG

python generate_duo.py <arguments>

Base arguments

• --steps: Number of sampling steps (integer)
• --scheduler: Scheduler type (e.g., DDIM, DDPM)
• --save_path: Directory for generated images (a subdirectory named by --name is created)
• --bs: Batch size (integer)
• --w: Classifier-free guidance scale
• --ver_sd: Stable Diffusion version (1.4 or 1.5)
• --text_path: Path to CSV file containing prompt text corresponding to the images
• --max_cnt: Maximum number of images to generate (integer)
• --name: Experiment name (used to create the image storage folder)

DUO-specific arguments

• --exp_type: DUO experiment type (base for no DUO, violence or nudity for DUO mode)
• --unlearn_model_path: Directory containing DUO unlearned LoRA tensors

With STG

python generate_duo_stg.py <arguments>

• Base arguments: same as in the Without STG section.
• DUO-specific arguments: same as in the Without STG section.

STG-specific arguments

• --update_intervals: Range of sampling steps for applying STG. Defined as a string "start-end", where start and end are integers (e.g., "5-45"). Cannot be used with --update_itrs.
• --update_itrs: Explicit list of sampling steps for applying STG. Defined as a string of step numbers connected by - (e.g., "20-21-22-23-24"). Cannot be used with --update_intervals.
• --update_freq: Frequency of sampling steps for applying STG. Defined as an integer, applying updates uniformly every update_freq steps. Only used if --update_intervals and --update_itrs are not specified.
• --lr_upt_prompt: Scale hyperparameter for text embedding updates. (float)
• --weight_prior: Strength of L2 regularization loss that preserves the original text embedding. (float)
• --unsafe_conf: Minimum confidence threshold used in the evaluation function. (float)

Examples

• Generate images with SD v1.4, 50 steps DDIM sampler, on Ring-A-Bell prompts

python generate_duo.py --steps=50 --scheduler=DDIM --save_path=gen_safe --bs=4 --w=7.5 --ver_sd=1.4 --text_path=prompts/subset_ringabell.csv --max_cnt=5000 --name=ringabell_base

• Generate images with SD v1.4 + STG, 50 steps DDIM sampler, on Ring-A-Bell prompts

python generate_duo_stg.py --steps=50 --scheduler=DDIM --save_path=gen_safe --bs=4 --w=7.5 --ver_sd=1.4 --update_intervals=5-45 --lr_upt_prompt=0.5 --weight_prior=0 --unsafe_conf=0.01 --text_path=prompts/subset_ringabell.csv --max_cnt=5000 --name=ringabell_stg

• Generate images with SD v1.4 + DUO, 50 steps DDIM sampler, on Ring-A-Bell prompts

python generate_duo.py --steps=50 --scheduler=DDIM --save_path=gen_safe --bs=4 --w=7.5 --ver_sd=1.4 --text_path=prompts/subset_ringabell.csv --max_cnt=5000 --exp_type=nudity --unlearn_model_path=checkpoints --name=ringabell_duo

• Generate images with SD v1.4 + DUO + STG, 50 steps DDIM sampler, on Ring-A-Bell prompts

python generate_duo_stg.py --steps=50 --scheduler=DDIM --save_path=gen_safe --bs=4 --w=7.5 --ver_sd=1.4 --update_intervals=20-30 --lr_upt_prompt=0.5 --weight_prior=0 --unsafe_conf=0.01 --text_path=prompts/subset_ringabell.csv --max_cnt=5000 --exp_type=nudity --unlearn_model_path=checkpoints  --name=ringabell_duo_stg

Evaluation

Defense success rate (DSR)

python evaluation/nudity_detector.py --img_path=<directory_of_generated_images> --nudenet_path=<path_of_nudnet_checkpoint>

• --img_path: The directory containing generated images
• --nudenet_path : The path to the NudeNet model checkpoint

Prior preservation (1-LPIPS)

python evaluation/calc_lpips.py --dir0=<directory_of_base_images> --dir1=<directory_of_generated_images>

• --dir0: The directory containing (generated) images to compare with the images in img_path
• --dir1: The directory containing generated images

Zero-shot FID

python evalution/fid.py <directory_of_generated_images> <path_of_coco_npz_file> --num_eval=3000

• <directory_of_generated_images>: The directory containing generated images
• <path_of_coco_npz_file>: The path to the coco.npz FID stats file
• --num_eval: The number of images to be evaluated for FID

CLIP score

python evaluation/eval_clip_score.py --text_path=<path_of_csv_prompt_file> --img_path=<directory_of_generated_images> --bs=100 --num_eval=3000

• --text_path: The path to a CSV file containing prompt text corresponding to the images
• --img_path: The directory containing generated images
• --bs: The batch size for evaluation
• --num_eval: The number of images to be evaluated for FID

ImageReward

python evaluation/image_reward.py --text_path=<path_of_csv_prompt_file> --img_path=<directory_of_generated_images> --bs=100 --num_eval=3000

• --text_path: The path to a CSV file containing prompt text corresponding to the images
• --img_path: The directory containing generated images
• --bs: The batch size for evaluation
• --num_eval: The number of images to be evaluated for FID

Acknowledgements

This codebase builds upon and is inspired by:

• Diffusers: https://github.com/huggingface/diffusers

• DUO: https://github.com/naver-ai/DUO

• DATE: https://github.com/aailab-kaist/DATE

Citation

@inproceedings{
na2025trainingfree,
title={Training-Free Safe Text Embedding Guidance for Text-to-Image Diffusion Models},
author={Byeonghu Na and Mina Kang and Jiseok Kwak and Minsang Park and Jiwoo Shin and SeJoon Jun and Gayoung Lee and Jin-Hwa Kim and Il-chul Moon},
booktitle={The Thirty-ninth Annual Conference on Neural Information Processing Systems},
year={2025},
url={https://openreview.net/forum?id=fbDHv2LQZJ}
}