Research map / Video / Video generation
Target-Aware Video Diffusion Models
Taeksoo Kim, Hanbyul Joo
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 cross-attention loss on the [TGT] token (lambda_attn 0.0 -> 0.1) (ablation) | Target-aware CogVideoX · Target-aware benchmark | Contact Score: 0.647 -> 0.878 | Improved |
| fine-tune CogVideoX on the curated dataset only (no mask or loss) | CogVideoX · Target-aware benchmark | Contact Score: 0.560 -> 0.638 | Improved |
| apply loss to V2T cross-attention rather than T2V (ablation) | Target-aware CogVideoX · Target-aware benchmark | Contact Score: 0.740 -> 0.878 | Improved |
| apply loss to V2T only rather than both T2V and V2T (ablation) | Target-aware CogVideoX · Target-aware benchmark | Contact Score: 0.860 -> 0.878 | Improved |
| choose loss blocks by semantic alignment instead of random or equally spaced (ablation) | Target-aware CogVideoX · Target-aware benchmark | Contact Score: 0.819 (random), 0.816 (equally spaced) -> 0.878 | Improved |
| increase lambda_attn beyond 0.1 (0.25, 0.5, 1.0) (ablation) | Target-aware CogVideoX · Target-aware benchmark | Contact Score: 0.878 -> 0.890, 0.888, 0.888 | No clear effect |
| dilate or erode the target mask (ablation) | Target-aware CogVideoX · Mask quality ablation | Contact Score: 0.896 -> 0.872 to 0.904 | No clear effect |
| replace mask with a circular mask (radius 15 or 30) (ablation) | Target-aware CogVideoX · Mask shape ablation | Contact Score: 0.896 -> 0.838 (r15), 0.888 (r30) | Mixed |
| use automatically generated masks (GPT-4o + Grounded-SAM) (ablation) | Target-aware CogVideoX · Target-aware benchmark (50 images) | Contact Score: 0.896 -> 0.864 | No clear effect |
| add captioner-generated descriptions to the general sentence in training prompts (ablation) | Target-aware CogVideoX · Target-aware benchmark | Contact Score: 0.705 -> 0.878 | Improved |
| selective loss on every 5th block in blocks 5-23 | Target-aware CogVideoX · VRAM usage | VRAM usage reduction: 71% reduction vs all blocks | Improved |
| add target-aware training vs attention modulation (Direct-a-video style) | CogVideoX · Target-aware benchmark | Contact Score: 0.560 -> 0.546 | Hurt |
Closest papers by meaning
- Emergent Temporal Correspondences from Video Diffusion Transformers (NeurIPS 2025 · similarity 0.88)
- Through-The-Mask: Mask-based Motion Trajectories for Image-to-Video Generation (CVPR 2025 · similarity 0.88)
- Track4Gen: Teaching Video Diffusion Models to Track Points Improves Video Generation (CVPR 2025 · similarity 0.88)
- Space-Time Diffusion Features for Zero-Shot Text-Driven Motion Transfer (CVPR 2024 · similarity 0.88)
- Video Diffusion Models are Training-free Motion Interpreter and Controller (NeurIPS 2024 · similarity 0.88)
- ShotAdapter: Text-to-Multi-Shot Video Generation with Diffusion Models (CVPR 2025 · 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.
