Position: What makes an image realistic?
Authors: Lucas Theis
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Theoretical | Here, we discuss the closely related problem of quantifying realism, that is, designing functions that can reliably tell realistic data from unrealistic data. This problem turns out to be significantly harder to solve and remains poorly understood, despite its prevalence in machine learning and recent breakthroughs in generative AI. Drawing on insights from algorithmic information theory, we discuss why this problem is challenging, why a good generative model alone is insufficient to solve it, and what a good solution would look like. In particular, we introduce the notion of a universal critic, which unlike adversarial critics does not require adversarial training. While universal critics are not immediately practical, they can serve both as a North Star for guiding practical implementations and as a tool for analyzing existing attempts to capture realism. |
| Researcher Affiliation | Industry | 1Google DeepMind, London, UK. Correspondence to: Lucas Theis <theis@google.com>. |
| Pseudocode | No | The paper does not contain any pseudocode or algorithm blocks. It primarily presents conceptual discussions and mathematical equations. |
| Open Source Code | No | The paper is a theoretical position paper and discusses concepts that are "uncomputable" in practice. It does not mention or provide any open-source code for its own described methodology or concepts. |
| Open Datasets | No | The paper mentions datasets like MNIST in illustrative examples to explain theoretical concepts, but it does not specify any dataset used for training or evaluation in experiments conducted by the authors, nor does it provide access information for such datasets related to its own work. |
| Dataset Splits | No | The paper is theoretical and does not describe any experiments conducted by the authors. Consequently, it does not provide details on training, validation, or test dataset splits. |
| Hardware Specification | No | The paper is theoretical and does not describe any experiments conducted by the authors. Therefore, it does not provide any hardware specifications. |
| Software Dependencies | No | The paper is theoretical and does not describe any experiments conducted by the authors. As such, it does not list any specific software dependencies with version numbers. |
| Experiment Setup | No | The paper is theoretical and does not describe any experiments conducted by the authors. Therefore, it does not provide details on experimental setup, hyperparameters, or system-level training settings. |