Research map / Image generation / Autoregressive and latent generation

Position: What makes an image realistic?

ICML 2024 · Image generation › Autoregressive and latent generation · approach: Generalizable fake image detection

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Position paper arguing that realism should be quantified via randomness deficiency from algorithmic information theory, introducing "universal critics" that need no adversarial training and relating existing metrics (divergences, typicality, SDS) to this view.

Problem

Designing reliable real-valued functions that score how realistic a sample (image, text, video) is, which can be optimized against without adversarial training, is poorly understood; common probability, typicality, and divergence-based notions fail.

Approach

Define realism as randomness deficiency U(x) = -log P(x) - K(x), approximated via a Solomonoff mixture; a batched version gives a sandwich bound to KL divergence and does not depend on the alternative distribution Q, so no adversarial critic is needed.

Key results

• Shows that the universal critic's batched average satisfies a sandwich bound: DKL[Q||P] - (1/B) log(1/pi_Q) <= E_Q[U^B]/B <= DKL[Q||P] (up to the stated terms), recovering KL in the large-batch limit.
• Argues that probability-based and weak-typicality realism measures fail on simple coin-toss and Gaussian examples.
• Reinterprets score distillation sampling's classifier-free guidance gradient as resembling a universal critic against a mixture of alternatives.

Surprising findings

• Maximizing probability or density can yield unrealistic samples (e.g., an all-ones coin-toss sequence is the most probable), and weak typicality is necessary but not sufficient for realism.

Limitations

• Universal critics depend on uncomputable Kolmogorov complexity and are not immediately practical.
• No empirical experiments; claims rest on theoretical argument and examples.
• Practical differentiable approximations remain open.

Experimental setup

Baselines: probability-based realism weak typicality adversarial critics maximum mean discrepancy Fréchet inception distance

Builds on: Solomonoff probability Neyman-Pearson lemma f-divergence variational bounds score distillation sampling input complexity (Serra et al. 2020)

Closest papers by meaning

• Be More Specific: Evaluating Object-centric Realism in Synthetic Images (CVPR 2025 · similarity 0.85)
• The Unmet Promise of Synthetic Training Images: Using Retrieved Real Images Performs Better (NeurIPS 2024 · similarity 0.84)
• Stay-Positive: A Case for Ignoring Real Image Features in Fake Image Detection (ICML 2025 · similarity 0.83)
• Generative Photography: Scene-Consistent Camera Control for Realistic Text-to-Image Synthesis (CVPR 2025 · similarity 0.83)
• Shadows Don't Lie and Lines Can't Bend! Generative Models don't know Projective Geometry...for now (CVPR 2024 · similarity 0.83)
• Looks Too Good To Be True: An Information-Theoretic Analysis of Hallucinations in Generative Restoration Models (NeurIPS 2024 · similarity 0.82)