Research map / Video / Video generation
Unity in Diversity: Video Editing via Gradient-Latent Purification
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 Gradient-Latent Purification (GLP), a training-free-style latent optimization framework that denoises DDS gradients and selects a stable final latent via SVD-based local coordinate systems for text-driven video editing on ZeroScope.
Problem
Latent-optimization-based text-driven video editing suffers from gradient noise from self-supervised objectives (e.g., DDS), causing blurriness and oversaturation, and from difficulty choosing the optimal stopping point, leading to under- or over-optimized edits.
Approach
Collects gradients over a short window and final-stage latents, builds per-frame local coordinate systems via SVD of low-variance axes, linearly recombines them with Gaussianity (skewness/kurtosis) regularization, and extends axes across adjacent frames for temporal smoothness.
Key results
- GLP reaches Text-Align 28.67 and Frame-Acc 0.84 on 40 DAVIS/WebVid videos, the best among compared methods (RAVE 28.35, 0.74).
- Ablation on DDS baseline: Text-Align rises from 27.09 to 28.67 and user-study text alignment from 2.78 to 3.22 after adding gradient and latent purification.
- Achieves motion temporal coherence -5.45 versus -5.88 for STEM-Inv and V-Quality 0.66 comparable to the best baselines.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add gradient purification to DDS (ablation) | DDS video editing with ZeroScope · Video editing | Text-Align: 27.09 -> 28.05 | Helped |
| Add gradient purification to DDS (ablation) | DDS video editing with ZeroScope · Video editing | Motion-TC: -5.71 -> -5.51 | Helped |
| Add latent purification after gradient purification (ablation) | DDS + gradient purification video editing with ZeroScope · Video editing | V-Quality: 0.63 -> 0.66 | Helped |
| Add latent purification after gradient purification (ablation) | DDS + gradient purification video editing with ZeroScope · Video editing | TC: 3.35 -> 3.47 | Helped |
| Use GLP rather than STEM-Inv | Video editing · Video editing | Semantic-TC: 0.94 -> 0.93 | Hurt |
| Use GLP rather than RAVE | Video editing · Video editing | V-Quality: 0.68 -> 0.66 | Hurt |
Surprising findings
- RAVE scores best on the video-level V-Quality (0.68) metric, above GLP's 0.66, so GLP is not uniformly best on quality metrics.
Limitations
- Struggles with fast-moving objects and motion blur, and with vehicle orientation alignment.
- Relies on hand-set hyperparameters (J, M, K_grad, K_latent) selected empirically.
- Evaluated on only 40 videos, with a 20-participant user study.
Experimental setup
Datasets: DAVIS WebVid
Benchmarks: DAVIS WebVid
Metrics: Text-Align Frame-Pick Frame-Acc Motion-TC Semantic-TC Object-TC V-Quality F-Quality user study
Baselines: TokenFlow FLATTEN RAVE STEM-Inv STDF DDS
Builds on: Delta Denoising Score ZeroScope DreamMotion classifier-free guidance
Closest papers by meaning
- FADE: Frequency-Aware Diffusion Model Factorization for Video Editing (CVPR 2025 · similarity 0.88)
- Señorita-2M: A High-Quality Instruction-based Dataset for General Video Editing by Video Specialists (NeurIPS 2025 · similarity 0.86)
- VideoDirector: Precise Video Editing via Text-to-Video Models (CVPR 2025 · similarity 0.86)
- Ground-A-Video: Zero-shot Grounded Video Editing using Text-to-image Diffusion Models (ICLR 2024 · similarity 0.86)
- CCEdit: Creative and Controllable Video Editing via Diffusion Models (CVPR 2024 · similarity 0.85)
- SplitFlow: Flow Decomposition for Inversion-Free Text-to-Image Editing (NeurIPS 2025 · similarity 0.85)
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.