Variational Diffusion Posterior Sampling with Midpoint Guidance
Authors: Badr MOUFAD, Yazid Janati el idrissi, Lisa Bedin, Alain Oliviero Durmus, randal douc, Eric Moulines, Jimmy Olsson
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We validate the proposed approach through extensive experiments on linear and nonlinear inverse problems, including challenging cases with latent diffusion models as priors. We then demonstrate its applicability to various modalities and its promising impact on public health by tackling cardiovascular disease diagnosis through the reconstruction of incomplete electrocardiograms. The code is publicly available at https://github.com/ yazidjanati/mgps. [...] 4 EXPERIMENTS We now evaluate our algorithm on three different problems and compare it with several competitors. We begin in benchmarking our method on toy Gaussian-mixture targets and image experiments with both pixel space and latent diffusion models. [...] For all tasks, we report the LPIPS (Zhang et al., 2018) between the reference image and the reconstruction, averaged over 50 images. For the five tasks tested on 1k images, we also report the FID in Table 7. |
| Researcher Affiliation | Academia | Badr Moufad ,1 Yazid Janati ,1 Lisa Bedin ,1 Alain Durmus1 Randal Douc2 Eric Moulines1 Jimmy Olsson3 1Ecole polytechnique 2Télécom Sud Paris 3KTH Royal Institute of Technology |
| Pseudocode | Yes | Algorithm 1 MIDPOINT GUIDANCE POSTERIOR SAMPLING [...] Algorithm 2 Warm Start [...] Algorithm 3 MGPS with warm start strategy |
| Open Source Code | Yes | The code is publicly available at https://github.com/ yazidjanati/mgps. [...] Our code for reproducing all the experiments is publicly available.1 1Code available at https://github.com/yazidjanati/mgps |
| Open Datasets | Yes | We test our algorithm on the 256 x 256 versions of the FFHQ (Karras et al., 2019) and Image Net (Deng et al., 2009) datasets using publicly available pre-trained DDMs. [...] We train a state-space diffusion model (Goel et al., 2022) to generate ECGs using the 20k training ECGs from the PTB-XL dataset (Wagner et al., 2020), and benchmark the posterior sampling algorithm on the 2k test ECGs; see appendix D.5. |
| Dataset Splits | Yes | Table 12: PTB-XL dataset description. Split All RBBB LBBB AF SB Train 17,403 432 428 1211 503 Val 2,183 55 54 151 64 Test 2,203 54 54 152 64 |
| Hardware Specification | No | The paper mentions "GPU memory requirements" and thanks "IHU-LIRYC for the computing power made available to us" but does not specify any particular GPU or CPU models, memory sizes, or other specific hardware configurations used for the experiments. |
| Software Dependencies | No | The paper mentions the use of the "Adam optimizer (Kingma & Ba, 2015)" and refers to architectures like "SSSDSA (Alcaraz & Strodthoff, 2022)", but it does not specify software dependencies with version numbers (e.g., Python 3.x, PyTorch 1.x, CUDA 1x.x). |
| Experiment Setup | Yes | We implement Algorithm 1 with ℓk = k/2 and use the Adam optimizer (Kingma & Ba, 2015) with a learning rate of 0.03 for optimization. The number of gradient steps is adjusted based on the complexity of the task: posterior sampling with the Image Net DDM prior or FFHQ LDM prior is more challenging and therefore requires additional gradient steps. Detailed hyperparameters are provided in Table 6. [...] Hyper-parameters are described in table 13. The selected model is the model of the last epoch. Training time is 30 hours for 2.5 seconds ECGs, 42 hours for 10 seconds ECGs. [...] We report in table 5 the balanced accuracy of diagnosing Left Bundle Branch Block (LBBB), Right Bundle Branch Block (RBBB), Atrial Fibrillation (AF), and Sinus Bradycardia (SB) using the downstream classifier proposed in Strodthoff et al. (2020) (see appendix D.5.1) applied to both ground-truth and to reconstructed samples from lead I. |