Research map / Image generation / Text-to-image diffusion
Guiding Diffusion Models with Semantically Degraded Conditions
Shilong Han, Yuming Zhang, Hongxia Wang
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 |
|---|---|---|---|
| replace null prompt with semantically degraded condition (CDG) | SD3 Medium · MS-COCO 2017 val | FID: 35.69 -> 34.05 | Improved |
| replace null prompt with CDG | Qwen-Image · MS-COCO 2017 val | FID: 42.45 -> 39.02 | Improved |
| replace null prompt with CDG | Qwen-Image · MS-COCO 2017 val | Aesthetic Score: 2.57 -> 2.54 | Hurt |
| replace null prompt with CDG | FLUX.1-dev · MS-COCO 2017 val | FID: 38.55 -> 37.11 | Improved |
| replace null prompt with CDG | SD3.5 · GenAI-Bench | Differentiation: 75.10 -> 78.74 | Improved |
| replace null prompt with CDG | SD3 · MS-COCO 2017 val | Aesthetic Score: 5.66 -> 5.70 (ICG reaches 5.79) | Mixed |
| add stratified degradation (content tokens first, then context-aggregating tokens) (ablation) | CDG · MS-COCO 2017 val, SD3, Rdeg=1.1 | VQA Score: 86.31 -> 92.21 | Improved |
| use WPR ranking instead of random ranking inside stratified degradation (ablation) | CDG · MS-COCO 2017 val, SD3, Rdeg=1.1 | FID: 34.17 -> 33.89 | No clear effect |
| use WPR ranking instead of reverse or random ranking in non-stratified degradation (ablation) | CDG · MS-COCO 2017 val, SD3, Rdeg=1.1 | FID: 50.73 (reverse), 47.02 (random) -> 35.06 | Improved |
| recompute WPR mask at every denoising step instead of once (ablation) | CDG · SD3, Rdeg=1.1 | VQA Score and time: VQA 92.21 -> 92.33, time 5.655 s -> 8.031 s | No clear effect |
| use one-time WPR mask computation | CDG · SD3 | time (s): 5.456 -> 5.655 (+3.6%) | Hurt |
| raise degradation ratio Rdeg beyond 1.0 (remove context-aggregating tokens) (ablation) | CDG · SD3, CLIP Score and Aesthetic Score | CLIP/Aesthetic | Mixed |
Closest papers by meaning
- Rethinking the Spatial Inconsistency in Classifier-Free Diffusion Guidance (CVPR 2024 · similarity 0.91)
- CFG++: Manifold-constrained Classifier Free Guidance for Diffusion Models (ICLR 2025 · similarity 0.90)
- Towards a Golden Classifier-Free Guidance Path via Foresight Fixed Point Iterations (NeurIPS 2025 · similarity 0.89)
- Adaptive Guidance: Training-free Acceleration of Conditional Diffusion Models (AAAI 2025 · similarity 0.89)
- Normalized Attention Guidance: Universal Negative Guidance for Diffusion Models (NeurIPS 2025 · similarity 0.88)
- Rethinking Training for De-biasing Text-to-Image Generation: Unlocking the Potential of Stable Diffusion (CVPR 2025 · similarity 0.88)
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.
