samplers makes posterior sampling with diffusion models almost plug-and-play.
Posterior sampling lets you tackle a broad class of inverse problems — such as denoising, inpainting, jpeg decompression, or deblurring — by combining two ingredients:
- A degradation operator that describes how your data were corrupted.
- A strong generative prior (any pretrained diffusion model, pixel or latent) that knows what clean signals look like.
Instead of training a new network for every task, you reuse an off-the-shelf model and draw samples from the posterior distribution p(x | y). The result: you can solve problems the model was never trained for, often with state-of-the-art quality and zero extra training.
- Bring-your-own model. Works with many diffusion models from Hugging Face, easily extendable.
- Operator + noise abstraction. Drop-in classes for blurs, masks, Gaussian or Poisson noise, … or write your own in a few lines.
- Pixel or latent space. Pairs naturally with frameworks such as Stable Diffusion.
- Typed, modular code. Clear abstract hierarchy.
- Lean dependency stack. PyTorch ≥ 2.0,
diffusers,torchvision— nothing exotic. - Growing algorithm zoo. State-of-the-art samplers already included, new ones coming out soon.
git clone https://github.com/thomashirtz/samplers
cd samplers
pip install -e .
from PIL import Image
from samplers.inverse_problem import InverseProblem
from samplers.networks import DDPMNetwork
from samplers.noise import GaussianNoise
from samplers.operators import IdentityOperator
from samplers.samplers import DPSSampler
from samplers.utils.image import pil_to_tensor, tensor_to_pil
# 1. Instantiate a pretrained prior
model_name = "google/ddpm-celebahq-256"
network = DDPMNetwork.from_pretrained(model_name)
# 2. Inverse problem construction
x_true = pil_to_tensor(Image.open("image.png"))
operator = IdentityOperator(x_shape=x_true.shape)
noise = GaussianNoise(sigma=0.05)
# For the demonstration; in practice inverse problems start from y = A(x) + noise
problem = InverseProblem.from_clean_data(x_true, noise, operator)
# 3. Sample from the posterior
sampler = DPSSampler(network)
x_hat = sampler(problem)
tensor_to_pil(x_hat).save("restored.png")Released under the BSD 3‑Clause License © 2025 Thomas Hirtz.
See LICENSE for the full text.