Diffusion Posterior Sampling for Linear Inverse Problem Solving: A Filtering Perspective

Abstract

Diffusion models have achieved tremendous success in generating high-dimensional data like images, videos and audio. These models provide powerful data priors that can solve linear inverse problems in zero shot through Bayesian posterior sampling. However, exact posterior sampling for diffusion models is intractable. Current solutions often hinge on approximations that are either computationally expensive or lack strong theoretical guarantees. In this work, we introduce an efficient diffusion sampling algorithm for linear inverse problems that is guaranteed to be asymptotically accurate. We reveal a link between Bayesian posterior sampling and Bayesian filtering in diffusion models, proving the former as a specific instance of the latter. Our method, termed filtering posterior sampling, leverages sequential Monte Carlo methods to solve the corresponding filtering problem. It seamlessly integrates with all Markovian diffusion samplers, requires no model re-training, and guarantees accurate samples from the Bayesian posterior as particle counts rise. Empirical tests demonstrate that our method generates better or comparable results than leading zero-shot diffusion posterior samplers on tasks like image inpainting, super-resolution, and motion deblur.

Prerequisites

• python 3.8

• pytorch >= 1.7.0

• CUDA >= 10.2

Here, the version of CUDA and pytorch need to be compatible (such as CUDA 10.2 + pytorch 1.7.0 or CUDA 11.3.1 + pytorch 1.11.0).

Getting started

1) Clone the repository

git clone https://github.com/ZehaoDou-official/FPS-SMC-2023

cd FPS-SMC-2023

2) Download pretrained checkpoint

In this section, we create a new folder named "models/".

mkdir models

Then, we download the checkpoint "ffhq_10m.pt" (for FFHQ dataset) and "imagenet256.pt" (for ImageNet dataset). After that, we paste these two pretraining score estimation models to "./models/".

Link for "ffhq_10m.pt" and "imagenet256.pt": link (cited from DPS-2022).
Another choice for ImageNet is "256x256_diffusion_uncond.pt": link

mv {DOWNLOAD_DIR}/ffhq_10m.pt ./models/
mv {DOWNLOAD_DIR}/imagenet256.pt ./models/

{DOWNLOAD_DIR} is the directory that you downloaded checkpoint to.

3) Download Dataset

Download the FFHQ validation set and paste these images into "./data/samples".
Download the ImageNet validation set and paste these images into "./data/val_images".

Currently, we put five example images in both of these two folders. Finally, we will need to download the entire validation set (which won't use too much storage).

4) Set environment

Local environment setting

Extended from DPS-2022, we use the external codes for motion-blurring and non-linear deblurring (to showcase only).

git clone https://github.com/VinAIResearch/blur-kernel-space-exploring bkse

git clone https://github.com/LeviBorodenko/motionblur motionblur

Install dependencies

conda create -n DPS python=3.8

conda activate DPS

pip install -r requirements.txt

pip install torch==1.11.0+cu113 torchvision==0.12.0+cu113 torchaudio==0.11.0 --extra-index-url https://download.pytorch.org/whl/cu113

5) Inference

python3 sample_condition.py \
--model_config=configs/model_config.yaml \
--diffusion_config=configs/diffusion_config.yaml \
--task_config={TASK-CONFIG} \
--c_rate=0.95 \
--particle_size=5;

Possible task configurations

Since this work aims to solve linear inverse problems by using Filtering Posterior Sampling, we only provide configurations for linear inverse tasks.

# Linear inverse problems (on FFHQ dataset)
- configs/super_resolution_config.yaml
- configs/gaussian_deblur_config.yaml
- configs/motion_deblur_config.yaml
- configs/inpainting_config.yaml

# Linear inverse problems (on ImageNet dataset)
- configs/super_resolution_imagenet_config.yaml
- configs/gaussian_deblur_imagenet_config.yaml
- configs/motion_deblur_imagenet_config.yaml
- configs/inpainting_imagenet_config.yaml

For the inpainting task, you can change the mask type (inpainting_box or inpainting_random) in the config file.

Example and Structure of task configurations

conditioning:
    method: ps # check candidates in guided_diffusion/condition_methods.py
    params:
        scale: 0.5

data:
    name: ffhq
    root: ./data/samples/

    # name: imagenet
    # root: ./data/val_images/

measurement:
    operator:
        name: inpainting # check candidates in guided_diffusion/measurements.py
    mask_opt:
        mask_type: box # It means that the mask for the inpainting task has the box-type.
        mask_len_range: !!python/tuple [128,129] # It means the size of the box-type mask is 128 x 128.
        image_size: 256

noise:
    name: gaussian
    sigma: 0.05 # The standard deviation of Gaussian noise.