Research map / Video / Video generation
FlexiClip: Locality-Preserving Free-Form Character Animation
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.
FlexiClip animates clipart by learning Bézier keypoint trajectories and Jacobian-field mesh deformations, adding temporal Jacobian corrections (pfODE), and a GFlowNet-inspired flow matching loss, all optimized with video SDS from ModelScope T2V.
Problem
Animating clipart with smooth, temporally coherent motion while preserving geometry; existing methods (AniClipart) show abrupt motion and distortions, and T2V/I2V models struggle with clipart style.
Approach
Decomposes total Jacobian into spatial Jacobian plus a learned temporal Jacobian correction whose evolution is modeled as a probability flow ODE, and adds a detailed-balance-inspired flow matching loss comparing score functions of corrected and uncorrected frames so temporal noise is progressively reduced.
Key results
- FlexiClip reaches CLIP score 0.9563 and X-CLIP 0.2102, above AniClipart (0.9401, 0.2075).
- Deformation Smoothness drops from 18.49 (AniClipart) to 13.76 and Animation Energy rises from 75.23 to 113.44.
- User study (30 participants) gives FlexiClip 94.90% identity preservation vs 83.63% for AniClipart.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use FlexiClip instead of AniClipart | Clipart animation generation · Clipart animations | CLIP score: 0.9401 -> 0.9563 | Helped |
| use FlexiClip instead of AniClipart | Clipart animation generation · Clipart animations | X-CLIP score: 0.2075 -> 0.2102 | Helped |
| use FlexiClip instead of AniClipart | Clipart animation generation · Clipart animations | GD: 50.98 -> 52.34 | Hurt |
| add temporal Jacobian with pfODE continuous dynamics (ablation) | FlexiClip · Clipart animations | MV: 23.00 -> 25.33 | Helped |
| add temporal Jacobian with pfODE continuous dynamics (ablation) | FlexiClip · Clipart animations | TC: 8.80 -> 8.14 | Helped |
| add temporal Jacobian with pfODE continuous dynamics (ablation) | FlexiClip · Clipart animations | GD: 51.50 -> 52.34 | Hurt |
| add temporal Jacobian with pfODE continuous dynamics (ablation) | FlexiClip · Clipart animations | DS: 14.00 -> 13.76 | Helped |
| add temporal Jacobian with pfODE continuous dynamics (ablation) | FlexiClip · Clipart animations | AE: 105.00 -> 113.44 | Helped |
| add flow matching loss (ablation) | FlexiClip · Clipart animations | TC: 8.40 -> 8.14 | Helped |
| add flow matching loss (ablation) | FlexiClip · Clipart animations | GD: 53.00 -> 52.34 | Helped |
| add flow matching loss (ablation) | FlexiClip · Clipart animations | AE: 95.00 -> 113.44 | Helped |
| increase λ from around 1 to up to 5 | FlexiClip · Text-conditioned motion generation | training steps to acceptable output | No clear effect |
Surprising findings
- Geometric Deviation is higher for FlexiClip (52.34) than AniClipart (50.98), which the authors attribute to the absence of ARAP deformation.
Limitations
- Geometric deviation is worse than AniClipart because ARAP deformation is not used.
- Evaluation relies on a small set of about 30 clipart images and a 30-participant user study.
- Gains over baselines on CLIP and X-CLIP scores are small in absolute terms.
Experimental setup
Datasets: AniClipart clipart images Freepik clipart
Benchmarks: AniClipart evaluation clipart set user study with 55 clipart images
Metrics: CLIP score X-CLIP score Motion Vibrancy Temporal Consistency Geometric Deviation Deformation Smoothness Animation Energy user selection %
Baselines: AniClipart Gal23 DynamiCrafter VideoCrafter2 I2VGen-XL ModelScope ToonCrafter AnimateLCM-I2V Pyramid Flow LTXVideo
Builds on: AniClipart video score distillation sampling probability flow ODE GFlowNet detailed balance DiffVG
Closest papers by meaning
- Animate Anyone: Consistent and Controllable Image-to-Video Synthesis for Character Animation (CVPR 2024 · similarity 0.87)
- Arbitrary Motion Style Transfer with Multi-condition Motion Latent Diffusion Model (CVPR 2024 · similarity 0.86)
- MagicAnimate: Temporally Consistent Human Image Animation using Diffusion Model (CVPR 2024 · similarity 0.86)
- Animate-X: Universal Character Image Animation with Enhanced Motion Representation (ICLR 2025 · similarity 0.85)
- EverAnimate: Minute-Scale Human Animation via Latent Flow Restoration (NeurIPS 2026 · similarity 0.85)
- AnimateDiff: Animate Your Personalized Text-to-Image Diffusion Models without Specific Tuning (ICLR 2024 · 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.