Research map / Video / Video generation
SIFT: Self-Imagination Fine-Tuning for Physically Plausible Motion in Video Diffusion Models
Ruoyu Wang, Jialun Liu, Huayang Huang, Haibin Huang, Jiepeng Wang, Chi Zhang, Xuelong Li, Yu Wu
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 self-imagination generation from pure noise instead of noisy real videos (ablation) | SIFT (Wan) · Camera-only (VLM, ablation) | Physical Commonsense: 3.55 -> 4.15 | Improved |
| add self-imagination generation from pure noise instead of noisy real videos (ablation) | SIFT (Wan) · Camera-only (VLM, ablation) | Semantic Adherence: 3.05 -> 3.85 | Improved |
| add alternating supervision between R3D and SlowFast instead of R3D only (ablation) | SIFT (Wan) · Camera-only (VLM, ablation) | Semantic Adherence: 3.65 -> 3.85 | Improved |
| add alternating supervision between R3D and SlowFast instead of SlowFast only (ablation) | SIFT (Wan) · Camera-only (VLM, ablation) | Physical Commonsense: 3.25 -> 4.15 | Improved |
| use single R3D classifier instead of alternating (object-only SA higher than full) (ablation) | SIFT (Wan) · Object-only (VLM, ablation) | Semantic Adherence: 3.50 -> 3.20 | Hurt |
| add progressive hard case replay (ablation) | SIFT (Wan) · Camera-only (VLM, ablation) | Semantic Adherence: 3.00 -> 3.85 | Improved |
| add progressive hard case replay (ablation) | SIFT (Wan) · Object-only (VLM, ablation) | Physical Commonsense: 3.70 -> 4.00 | Improved |
| fine-tune with SFT on 4,000 motion-decoupled videos | Wan · Human eval object-only | Physical Commonsense: 2.75 -> 2.43 | Hurt |
| fine-tune with SFT on 4,000 motion-decoupled videos | Wan · Camera-only (VLM) | Physical Commonsense: 4.06 -> 4.73 | Improved |
| replace SFT with SIFT | Wan · Camera-only (VLM) | Physical Commonsense: 4.73 -> 4.93 | Improved |
| replace VideoREPA with SIFT | CogVideoX · Object-only (VLM) | Physical Commonsense: 3.50 -> 4.25 | Improved |
| add SIFT on multi-object motion | Wan · Multi-object (50 prompts) | Physical Commonsense: 3.92 -> 4.18 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Wan-Move — Wan-Move: Motion-controllable Video Generation via Latent Trajectory Guidance (NeurIPS 2025) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- DIVER: Disentangling Camera–Object and Active–Passive Motion for Video Generation (ECCV 2026 · similarity 0.90)
- LaMo: Self-Supervised Latent Motion Priors for Physical Realism in Video Generation (NeurIPS 2026 · similarity 0.89)
- Video Diffusion Models are Training-free Motion Interpreter and Controller (NeurIPS 2024 · similarity 0.89)
- FlowMo: Variance-Based Flow Guidance for Coherent Motion in Video Generation (NeurIPS 2025 · similarity 0.89)
- VideoJAM: Joint Appearance-Motion Representations for Enhanced Motion Generation in Video Models (ICML 2025 · similarity 0.89)
- Motion Consistency Model: Accelerating Video Diffusion with Disentangled Motion-Appearance Distillation (NeurIPS 2024 · similarity 0.89)
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.
