Official code for Slowly Annealed Langevin Dynamics: Theory and Applications to Training-Free Guided Generation.

Project page: https://sites.google.com/site/atsushinitanda/va-sald

Paper: https://arxiv.org/abs/2605.07950

VA-SALD is a training-free guided generation framework for pretrained generative models whose marginal path is induced by an Itô diffusion. The implementation principle is to adapt the VA-SALD drift to the backbone's own forward process, rather than treating every sampler as the same black box. This repository currently exposes two reproducible examples:

• sald_Gaussian_Moon/: synthetic VP-diffusion experiments comparing SALD, VA-SALD, and DOIT on guided Gaussian mixtures.
• sald_FM_SD35/: VA-SALD for the flow-matching Itô formulation of Stable Diffusion 3.5 Medium with zeroth-order black-box guidance.

The synthetic experiments are fully reproducible on CPU or GPU and include the main paper panels.

For the text-to-image experiment, sald_FM_SD35 implements the flow-matching version of VA-SALD for SD3.5-M and supports Aesthetic, PickScore, and CLIP-based guidance through zeroth-order reward gradients.

Clone the repository:

git clone https://github.com/anitan0925/sald.git
cd sald

Run the lightweight synthetic experiments:

cd sald_Gaussian_Moon
python3 -m venv .venv
source .venv/bin/activate
python3 -m pip install --upgrade pip
python3 -m pip install -r requirements.txt
python3 run_reproduce.py --cpu --smoke

If python3 -m venv is unavailable on a Debian/Ubuntu machine, install the system package python3-venv first, or create an equivalent Conda environment and install sald_Gaussian_Moon/requirements.txt.

Run the full synthetic reproduction on a GPU:

python3 run_reproduce.py --gpu 0

Set up the SD3.5 flow-matching experiment:

cd ../sald_FM_SD35
conda env create -f environment.yaml
conda activate fsd
export PYTHONPATH=$(pwd)
CUDA_VISIBLE_DEVICES=0 accelerate launch --num_processes=1 --main_process_port=0 \
  trainer/VA_SALD_Guidance_zerothorder.py \
  --config config/VA_SALD_Guidance_zerothorder.py:base

The SD3.5 experiment downloads pretrained Hugging Face models and therefore requires the usual Hugging Face access setup for stabilityai/stable-diffusion-3.5-medium.

VA-SALD should be implemented using the pretrained model's own forward Itô process:

For a guide f_t and a slowdown schedule t(s), the continuous-time VA-SALD dynamics are

Equivalently,

The exact implementation depends on what the backbone predicts: noise, score, velocity, denoised sample, or flow-matching velocity. We include an AI-assistant prompt template in docs/AI_ASSISTANT_PLUGIN_PROMPT.md that asks an assistant to derive and implement the correct VA-SALD update for a new diffusion-based model.

sald_Gaussian_Moon/   Synthetic VP-diffusion experiments and paper figures.
sald_FM_SD35/         Flow-matching VA-SALD for Stable Diffusion 3.5 Medium.
docs/                 Plug-in adaptation prompt and implementation notes.
LICENSE               MIT license.
CITATION.bib          BibTeX citation.

@article{nitanda2026SALD,
  title = {Slowly Annealed Langevin Dynamics: Theory and Applications to Training-Free Guided Generation},
  author={Nitanda, Atsushi and Bu Dake and Lyu, Yueming and Veeravalli, Tanya},
  journal={arXiv:2605.07950},
  year={2026}
}

The synthetic experiments include a DOIT-style baseline adapted from the https://github.com/liamyzq/Doob_training_free_adaptation. The SD3.5 flow-matching experiment builds on Hugging Face Diffusers and pretrained Stability AI models.