Research map / Image generation / Text-to-image diffusion
Mutual Information Divergence: A Unified Metric for Multimodal Generative Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Gaussian cross-mutual information over CLIP features as the evaluation metric | multimodal generation evaluation · LAFITE generated Likert judgments | Kendall τ_c: 73.7 -> 87.3 | Helped |
| Use Gaussian cross-mutual information over CLIP features as the evaluation metric | multimodal generation evaluation · AMT human Likert judgments | Kendall τ_c: 11.1 -> 12.4 | Helped |
| Use Gaussian cross-mutual information over CLIP features as the evaluation metric | multimodal generation evaluation · VQ-Diffusion generated Likert judgments | Kendall τ_c: 78.0 -> 82.0 | Helped |
| Use MID instead of InfoNCE for foiled-caption visual reasoning evaluation | CLIP ViT-L/14 · Visual reasoning: count | accuracy: 0.230 -> 0.443 | Helped |
| Use MID instead of RefCLIP-S | image caption evaluation · Flickr8K-Expert | Kendall τ_c: 53.0 -> 54.9 | Helped |
| Use MID instead of RefCLIP-S | image caption evaluation · Flickr8K-CF | Kendall τ_b: 36.4 -> 37.3 | Helped |
| Use MID instead of RefCLIP-S | image caption evaluation · Pascal-50S mean | accuracy: 83.1 -> 85.2 | Helped |
| Use MID instead of RefCLIP-S | image caption evaluation · FOIL hallucination detection, 1 reference | accuracy: 91.0 -> 90.5 | Hurt |
| Use MID instead of RefCLIP-S | image caption evaluation · FOIL hallucination detection, 4 references | accuracy: 92.6 -> 90.5 | Hurt |
| Add a reference-caption cosine term to MID (RefMID) | MID on FOIL hallucination detection · FOIL hallucination detection | accuracy: 90.5 -> 92.4; 90.5 -> 93.7 | Helped |
| Use CLIP ViT-L/14 rather than CLIP ViT-B/32 | MID · LAFITE generated Likert judgments | Kendall τ_c: 74.6 -> 87.3 | Helped |
| Use CLIP ViT-L/14 rather than CLIP ViT-B/32 | MID · AMT human Likert judgments | Kendall τ_c: 11.9 -> 12.4 | Helped |
Closest papers by meaning
- Diffusion Feedback Helps CLIP See Better (ICLR 2025 · similarity 0.87)
- Information Theoretic Text-to-Image Alignment (ICLR 2025 · similarity 0.87)
- CapsFusion: Rethinking Image-Text Data at Scale (CVPR 2024 · similarity 0.86)
- SILMM: Self-Improving Large Multimodal Models for Compositional Text-to-Image Generation (CVPR 2025 · similarity 0.86)
- The Narrow Gate: Localized Image-Text Communication in Native Multimodal Models (NeurIPS 2025 · similarity 0.86)
- Sieve: Multimodal Dataset Pruning using Image Captioning Models (CVPR 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.