Research map / Image generation / Text-to-image diffusion
Anchoring and Steering Diffusion: Enhancing the Faithfulness of Text-to-Image Generation at Inference Time
Xinyi Wang, Yuyang Huang, Yalin Su, Pengcheng Luan, Tao Zhang, FEIMING WEI, Wenxian Yu
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 |
|---|---|---|---|
| add Semantic Anchoring (ablation) | SDXL · GenEval | overall accuracy: 55.933 -> 60.731 | Improved |
| add Reflective Steering (ablation) | SDXL · GenEval | overall accuracy: 55.933 -> 59.868 | Improved |
| add Reflective Steering to Semantic Anchoring (ablation) | SDXL · GenEval | overall accuracy: 60.731 -> 63.393 | Improved |
| add Semantic Anchoring (ablation) | HunyuanDiT-v1.2 · GenEval | overall accuracy: 68.342 -> 70.395 | Improved |
| add Reflective Steering (ablation) | HunyuanDiT-v1.2 · GenEval | overall accuracy: 68.342 -> 71.728 | Improved |
| add Reflective Steering to Semantic Anchoring (ablation) | HunyuanDiT-v1.2 · GenEval | overall accuracy: 70.395 -> 72.470 | Improved |
| add LP-SDS refinement (N_lp=400) to CLIP-only prior (ablation) | AnchorSteer on SDXL · GenEval | overall accuracy: 60.768 -> 63.393 | Improved |
| increase LP-SDS steps from 400 to 600 (ablation) | AnchorSteer on SDXL · GenEval | overall accuracy: 63.393 -> 62.512 | Hurt |
| increase inspection stride from 1 to 5 (ablation) | AnchorSteer on SDXL · GenEval | overall accuracy: 63.393 -> 59.950 | Hurt |
| use a 2B VLM instead of Qwen-VL-Chat for diagnosis (ablation) | AnchorSteer on SDXL · GenEval | overall accuracy: 63.393 -> 62.073 | Mixed |
| apply AnchorSteer on top of standard sampling | SDXL · GenEval image quality | ImageReward: 0.4891 -> 0.8781 | Improved |
| add Semantic Anchoring (ablation) | SDXL · GenEval sample diversity | 1-LPIPS (lower is more diverse): 0.3860 -> 0.4406 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- InitNO — InitNO: Boosting Text-to-Image Diffusion Models via Initial Noise Optimization (CVPR 2024) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Rethinking Training for De-biasing Text-to-Image Generation: Unlocking the Potential of Stable Diffusion (CVPR 2025 · similarity 0.90)
- InitNO: Boosting Text-to-Image Diffusion Models via Initial Noise Optimization (CVPR 2024 · similarity 0.90)
- Noise Diffusion for Enhancing Semantic Faithfulness in Text-to-Image Synthesis (CVPR 2025 · similarity 0.90)
- Towards Transformer-Based Aligned Generation with Self-Coherence Guidance (CVPR 2025 · similarity 0.90)
- Diff-Aid: Inference-time Adaptive Interaction Denoising for Rectified Text-to-Image Generation (NeurIPS 2026 · similarity 0.90)
- Image Regeneration: Evaluating Text-to-Image Model via Generating Identical Image with Multimodal Large Language Models (AAAI 2025 · similarity 0.90)
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.
