Research map / Image generation / Text-to-image diffusion
AlignVid: Taming Visual Dominance via Training-Free Attention Modulation in Text-guided Image-to-Video Generation
Yexin Liu, Wen-Jie Shu, Zile Huang, Haoze Zheng, Yueze Wang, Manyuan Zhang, Jinjing Zhu, Ser-Nam Lim, Harry Yang
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 AlignVid attention scaling | Wan2.1 · OmitI2V addition accuracy | accuracy: 71.75 -> 79.54 | Improved |
| add AlignVid attention scaling | FramePack · OmitI2V dynamic degree | dynamic degree: 20.05 -> 28.53 | Improved |
| add AlignVid attention scaling | Wan2.1 · OmitI2V aesthetic quality | aesthetic quality: 63.12 -> 61.63 | Hurt |
| use energy-based modulation instead of scalar scaling (ablation) | AlignVid on Wan2.1 · OmitI2V addition accuracy | accuracy: 80.76 -> 75.65 | Hurt |
| apply ASM to all denoising steps instead of early steps (ablation) | AlignVid on FramePack · OmitI2V | aesthetic quality: 63.41 -> 61.56 | Hurt |
| apply ASM only in middle denoising steps (ablation) | AlignVid on FramePack · OmitI2V modification accuracy | accuracy: 64.99 -> 62.71 | Hurt |
| restrict ASM to foreground-focus blocks instead of background-focus blocks (ablation) | AlignVid on Wan2.1 · OmitI2V modification accuracy | accuracy: 71.99 -> 77.20 | Improved |
| restrict ASM to first half of blocks instead of last half (ablation) | AlignVid on Wan2.1 · OmitI2V modification accuracy | accuracy: 73.68 -> 76.55 | Improved |
| scale only image keys in FramePack (Key-image) instead of both image and text keys (ablation) | AlignVid on FramePack · OmitI2V aesthetic quality | aesthetic quality: 63.57 -> 24.63 | Hurt |
| add AlignVid on top of CFG=5 (ablation) | Wan2.1 · OmitI2V modification accuracy | accuracy: 72.35 -> 77.20 | Improved |
| add AlignVid to OmniGen2 | OmniGen2 · GenEval Counting | score: 0.67 -> 0.52(-0.15) | Hurt |
| add AlignVid to Wan2.1-T2V-1.3B | Wan2.1-T2V-1.3B · VBench Multiple Objects | score: 53.35 -> 66.46(+13.11) | Improved |
Closest papers by meaning
- Focus-N-Fix: Region-Aware Fine-Tuning for Text-to-Image Generation (CVPR 2025 · similarity 0.89)
- Inference-Time Text-to-Video Alignment with Diffusion Latent Beam Search (NeurIPS 2025 · similarity 0.88)
- Devil is in the Detail: Towards Injecting Fine Details of Image Prompt in Image Generation via Conflict-free Guidance and Stratified Attention (CVPR 2025 · similarity 0.88)
- PromptCoT: Align Prompt Distribution via Adapted Chain-of-Thought (CVPR 2024 · similarity 0.87)
- PAROAttention: Pattern-Aware ReOrdering for Efficient Sparse and Quantized Attention in Visual Generation Models (NeurIPS 2025 · similarity 0.87)
- When Language Overrules: Revealing Text Dominance in Multimodal Large Language Models (NeurIPS 2026 · similarity 0.87)
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.
