Conditional Variational Diffusion Models
Authors: Gabriel Della Maggiora, Luis Alberto Croquevielle, Nikita Deshpande, Harry Horsley, Thomas Heinis, Artur Yakimovich
ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | This approach is tested in two unrelated inverse problems: super-resolution microscopy and quantitative phase imaging, yielding comparable or superior results to previous methods and fine-tuned diffusion models. We assess the performance of our model on three distinct benchmarks. |
| Researcher Affiliation | Academia | 1 Center for Advanced Systems Understanding (CASUS), G orlitz, Germany 2 Helmholtz-Zentrum Dresden-Rossendorf e. V. (HZDR), Dresden, Germany 3 Department of Computing, Imperial College London, London, United Kingdom 4 Bladder Infection and Immunity Group (BIIG), UCL Centre for Kidney and Bladder Health, Division of Medicine, University College London, Royal Free Hospital Campus, London, United Kingdom 5 Institute of Computer Science, University of Wrocław, Wrocław, Poland |
| Pseudocode | Yes | Figure 5 presents a high-level overview of the training algorithm and the inference process for CVDM. Algorithm 1 describes the training of the denoiser using a learnable schedule, while Algorithm 2 demonstrates the inference process and how to use the learned schedule during this procedure. |
| Open Source Code | Yes | 1The code is available on https://github.com/casus/cvdm. |
| Open Datasets | Yes | In this context, we utilize the Bio SR dataset (Qiao et al., 2021). We train the model with a synthetic dataset created by using Image Net to simulate Id and I d from grayscale images. We estimate the phase of images using the HCOCO dataset (Cong et al., 2019). |
| Dataset Splits | No | The paper mentions training data and test splits for some datasets (e.g., 'test split of HCOCO'), but it does not provide specific percentages or sample counts for validation splits, nor does it refer to predefined validation splits with citations. |
| Hardware Specification | No | The paper does not provide specific details about the hardware (e.g., GPU/CPU models, memory specifications) used to run the experiments. |
| Software Dependencies | No | The paper discusses the use of certain architectures (e.g., U-Net) and frameworks (e.g., Kingma et al. (2023)'s framework) but does not provide specific version numbers for any software dependencies or libraries used for implementation. |
| Experiment Setup | Yes | All models are trained for 400,000 iterations. For CDDPM, fine-tuning results in optimal performance for a linear schedule ranging from 0.0001 to 0.03. The model incorporates 5 scales, doubling the number of filters at each scale while concurrently halving the resolution. Each block consists of two sequences: convolution, activation, and instance normalization. Upscaling is achieved using transposed convolutions to ensure differentiability. Softplus activations are used throughout the architecture. |