Research map / Video / Video generation
CAMEL: CAusal Motion Enhancement Tailored for Lifting Text-driven Video Editing
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CAMEL, which disentangles motion from appearance in text-driven video editing using learnable motion prompts and a causal motion filter with motion-enhanced attention applied to high-frequency components of a Stable Diffusion UNet.
Problem
Text-driven video editing must keep flicker-free content consistency and preserve original motion patterns, but existing methods couple motion and appearance, so they either overfit appearance and miss motion or capture motion at the expense of creative text generalization.
Approach
Decompose temporal latent features into high- and low-frequency components with a learnable causal filter (generalizing the Haar wavelet). High-frequency components (motion) are enhanced via cross-attention to optimized motion prompts, then recombined with low-frequency components (appearance), while only the query projections and the new modules are fine-tuned.
Key results
- On TGVE-DAVIS, CAMEL reaches UMTScore 36.52 (object), 36.60 (background), 34.08 (style), 33.92 (multiple), beating Tune-A-Video by 1.5/1.62/0.77/1.16 points.
- On TGVE-Videvo, CAMEL reaches 37.03/37.54/35.26/37.36 UMTScore, the best of the compared methods across all four edit types.
- Ablation on TGVE-DAVIS: the causal motion filter plus motion prompts (row 4) gives the highest alignment, versus 35.03 average-style baseline with Haar wavelet alone (row 1).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CAMEL instead of Tune-A-Video | Stable Diffusion v1.4 video editing · TGVE-DAVIS Object | UMTScore: 35.02 -> 36.52 | Helped |
| Use CAMEL instead of Tune-A-Video | Stable Diffusion v1.4 video editing · TGVE-Videvo Background | UMTScore: 36.15 -> 37.54 | Helped |
| Use CAMEL instead of the strongest listed baseline | Stable Diffusion v1.4 video editing · TGVE-DAVIS Object | frame consistency: 92.19 -> 93.35 | Helped |
| Add motion prompts (ablation) | CAMEL on TGVE-DAVIS · Object editing | UMTScore: 36.03 -> 36.52 | Helped |
| Add motion prompts (ablation) | CAMEL on TGVE-DAVIS · Background changes | UMTScore: 35.98 -> 36.60 | Helped |
| Add the causal motion filter (ablation) | CAMEL on TGVE-DAVIS · Object editing | UMTScore: 35.31 -> 36.52 | Helped |
| Add the causal motion filter (ablation) | CAMEL on TGVE-DAVIS · Background changes | UMTScore: 35.24 -> 36.60 | Helped |
| Use a causal motion filter instead of Haar Wavelet (ablation) | CAMEL without motion prompts on TGVE-DAVIS · Object editing | UMTScore: 35.03 -> 36.03 | Helped |
Limitations
- Evaluation limited to 53 videos from the LOVEU-TGVE competition.
- Haar wavelet with a two-element window is too narrow for large motion changes, motivating the learned filter.
- Relies on structure guidance via DDIM inversion of the source video.
Experimental setup
Datasets: LOVEU-TGVE competition videos (DAVIS and Videvo subsets)
Benchmarks: LOVEU-TGVE (TGVE-DAVIS, TGVE-Videvo)
Metrics: UMTScore video-text alignment CLIP frame consistency
Baselines: Tune-A-Video Text2Video-Zero ControlVideo
Builds on: Tune-A-Video Stable Diffusion DDIM inversion Haar wavelet transform
Closest papers by meaning
- MotionEditor: Editing Video Motion via Content-Aware Diffusion (CVPR 2024 · similarity 0.87)
- Enhancing Motion in Text-to-Video Generation with Decomposed Encoding and Conditioning (NeurIPS 2024 · similarity 0.87)
- SteerVTE: Seamless Video Text Editing with Style and Glyph Control (NeurIPS 2026 · similarity 0.86)
- Textualize Visual Prompt for Image Editing via Diffusion Bridge (AAAI 2025 · similarity 0.86)
- VMC: Video Motion Customization using Temporal Attention Adaption for Text-to-Video Diffusion Models (CVPR 2024 · similarity 0.86)
- FADE: Frequency-Aware Diffusion Model Factorization for Video Editing (CVPR 2025 · 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.