metadata
license: apache-2.0
library_name: diffusers
pipeline_tag: text-to-image
SafeDiffusion-R1: Online Reward Steering for Safe Diffusion Post-Training
SafeDiffusion-R1 is a safety post-training framework for Stable Diffusion based on Group Relative Policy Optimization (GRPO). It uses a closed-form, CLIP-based steering reward to bake safety priors directly into the UNet weights, eliminating the need for separately trained safety classifiers or inference-time interventions.
Project Page | GitHub | Paper
Model Variants
The models are released as full Diffusers pipelines in different subfolders:
| Subfolder | Description |
|---|---|
scaled |
Main paper checkpoint. Best balance of safety and utility (Default). |
compact |
Optimized for lowest MMA-Diffusion ASR (adversarial robustness). |
empty-positive |
Ablation variant trained without safe anchors. |
Sample Usage
You can load and use the model variants using the diffusers library. Since the repository uses subfolders for different variants, we recommend using snapshot_download to load the specific version you need.
from huggingface_hub import snapshot_download
from diffusers import StableDiffusionPipeline
import os, torch
# Download the variant you want (e.g., "scaled")
local_root = snapshot_download(
"ItsMaxNorm/SafeDiffusion-R1",
allow_patterns="scaled/*", # or "compact/*" / "empty-positive/*"
)
# Load the pipeline
pipe = StableDiffusionPipeline.from_pretrained(
os.path.join(local_root, "scaled"),
torch_dtype=torch.float16,
).to("cuda")
# Generate an image
prompt = "a photo of a cat sleeping on a couch"
img = pipe(prompt).images[0]
img.save("out.png")
Citation
@misc{kumar2026safediffusionr1,
title={SafeDiffusion-R1: Online Reward Steering for Safe Diffusion Post-Training},
author={Komal Kumar and Ankan Deria and Abhishek Basu and Fahad Shamshad and Hisham Cholakkal and Karthik Nandakumar},
year={2026},
eprint={2605.18719},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2605.18719},
}