Research map / Image generation / Fast diffusion sampling
Conjuring Semantic Similarity
Tian Yu Liu, Stefano Soatto
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 trajectory-wide Jeffreys divergence instead of only initial-step noise prediction (ablation) | Conjuring Semantic Similarity · STS/SICK-R average | Spearman x100: 53.0 -> 65.4 | Improved |
| use trajectory-wide divergence instead of only final-step prediction (ablation) | Conjuring Semantic Similarity · STS/SICK-R average | Spearman x100: 56.9 -> 65.4 | Improved |
| compare the SDE scores instead of the generated output images (ablation) | Conjuring Semantic Similarity · STS/SICK-R average | Spearman x100: 51.3 -> 65.4 | Improved |
| symmetrize KL into Jeffreys divergence (ablation) | Conjuring Semantic Similarity · STS/SICK-R average | Spearman x100: 64.0 -> 65.4 | Improved |
| uniform prior over all timesteps instead of subset or single timestep (ablation) | Conjuring Semantic Similarity · STS-B | Spearman x100 | Improved |
| increase number of denoising steps T beyond 10 (ablation) | Conjuring Semantic Similarity · STS-B | Spearman x100: 70.1 (T=5), 70.3 (T=10), 69.5 (T=50) | No clear effect |
| increase Monte-Carlo steps k from 1 to 5 (ablation) | Conjuring Semantic Similarity · STS-B | Spearman x100: deviation of +/- 0.77 | No clear effect |
| swap Stable Diffusion v1.4 for SDXL or SD3 Medium (ablation) | Conjuring Semantic Similarity · STS-B | Spearman x100 | No clear effect |
| learn reverse diffusion for image generation on top of text encoder | Stable Diffusion text encoder · SimLex-999 verbs | Spearman x100: 10.3 -> -14.3 | Hurt |
| learn reverse diffusion on top of text encoder | Stable Diffusion text encoder · RG65 nouns | Spearman x100: 77.5 -> 60.2 | Hurt |
Closest papers by meaning
- The Hidden Language of Diffusion Models (ICLR 2024 · similarity 0.89)
- Evaluating Semantic Variation in Text-to-Image Synthesis: A Causal Perspective (ICLR 2025 · similarity 0.87)
- Interpretable Measures of Conceptual Similarity by Complexity-Constrained Descriptive Auto-Encoding (CVPR 2024 · similarity 0.87)
- CONFORM: Contrast is All You Need for High-Fidelity Text-to-Image Diffusion Models (CVPR 2024 · similarity 0.86)
- Image Regeneration: Evaluating Text-to-Image Model via Generating Identical Image with Multimodal Large Language Models (AAAI 2025 · similarity 0.86)
- Localizing and Editing Knowledge In Text-to-Image Generative Models (ICLR 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.
