Research map / Video / Video generation
LaMD: Latent Motion Diffusion for Image-Conditional Video Generation
Yaosi Hu, Zhenzhong Chen, Chong Luo
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 |
|---|---|---|---|
| raise motion channel size from d=1 to d=3 (ablation) | MCD-VAE · BAIR reconstruction | FVD: 41.91 -> 36.05 | Improved |
| raise motion channel size from d=3 to d=4 (ablation) | MCD-VAE · BAIR reconstruction | FVD: 36.05 -> 38.84 | Hurt |
| keep a temporal dimension in latent motion (d_t=2, d=2) instead of removing it (ablation) | MCD-VAE · BAIR reconstruction | FVD: 36.05 -> 37.93 | No clear effect |
| use larger latent space (beta e-5, d=4) instead of smaller (beta e-2, d=3) (ablation) | LaMD · Landscape generation | FVD: 100.7 -> 127.1 | Hurt |
| use larger latent space (beta e-5, d=4) instead of smaller (beta e-2, d=3) (ablation) | MCD-VAE · Landscape reconstruction | FVD: 58.59 -> 37.35 | Improved |
| MCD-VAE motion-decomposed autoencoder instead of VQGAN | video autoencoder · BAIR reconstruction | FVD: 42.80 -> 36.05 | Improved |
| generate latent motion with 2D-UNet (LaMD) instead of flow sequence with 3D-UNet (LFDM) | latent motion diffusion · 128-frame 128x128 sampling | sampling time (s): 40.1s -> 10.7s | Improved |
| latent motion diffusion instead of latent video diffusion | LVDM · 128-frame 128x128 sampling | sampling time (s): 26.6s -> 10.7s | Improved |
| LaMD compared with open-domain DynamiCrafter on dynamics | LaMD · Landscape motion quality | dynamic degree: 94.60% vs 55.13% | Hurt |
| LaMD instead of cINN on motion smoothness | LaMD · Landscape motion quality | motion smoothness: 99.41% -> 99.47% | 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.
- MOSO — MOSO: Decomposing MOtion, Scene and Object for Video Prediction (CVPR 2023) · better on 1 of 1 shared result
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
- Efficient Video Diffusion Models via Content-Frame Motion-Latent Decomposition (ICLR 2024 · similarity 0.92)
- LaMo: Self-Supervised Latent Motion Priors for Physical Realism in Video Generation (NeurIPS 2026 · similarity 0.90)
- Extrapolating and Decoupling Image-to-Video Generation Models: Motion Modeling is Easier Than You Think (CVPR 2025 · similarity 0.89)
- Decouple Content and Motion for Conditional Image-to-Video Generation (AAAI 2024 · similarity 0.88)
- Identifying and Solving Conditional Image Leakage in Image-to-Video Diffusion Model (NeurIPS 2024 · similarity 0.88)
- SF-V: Single Forward Video Generation Model (NeurIPS 2024 · similarity 0.88)
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.
