README.md

metadata

title: RADS — Steering Away from Memorization in Text-to-Image Diffusion
emoji: 🧭
colorFrom: green
colorTo: purple
sdk: gradio
sdk_version: 6.32.0
app_file: app.py
python_version: '3.12'
startup_duration_timeout: 30m
short_description: Steer SD v1.4 away from memorization at inference time

RADS: Reachability-Constrained Reinforcement Learning for Diffusion Steering

Demo for the paper "Steering Away from Memorization: Reachability-Constrained Reinforcement Learning for Text-to-Image Diffusion" (HF paper page, project page).

What this demo does

RADS is an inference-time steering controller for text-to-image diffusion. A Soft Actor-Critic policy — trained offline against a reachability-based safety critic — perturbs the prompt's caption embedding in a learned 64-d latent space at every denoising step, keeping generation inside a safe (Non-Memorizing) region of the diffusion trajectory. The diffusion backbone (Stable Diffusion v1.4, DDIM, T=50, g=7.5 — the paper's evaluation setting) is never modified.

For the same prompt and the same initial latent, this demo generates:

• Left: the plain SD v1.4 baseline (steering strength 0)
• Right: the RADS-steered image (steering strength 1, the paper default)

Both images come from an identical seed, so every visible difference is caused by steering alone. Example prompts are memorization-prone prompts from the paper's evaluation sets (Webster et al. 2023 and MemBench).

Checkpoints

SAC agent checkpoints (encoder + actor, training seeds 41–45) and the CLIP caption-embedding autoencoder are loaded from S-Karnik/rads-model-checkpoints. The official implementation lives at S-Karnik/rads-paper-repo (MIT).

Citation

@inproceedings{karnik2026rads,
  title   = {Steering Away from Memorization: Reachability-Constrained
             Reinforcement Learning for Text-to-Image Diffusion},
  author  = {Karnik, Shreyas and others},
  year    = {2026}
}