Research map / Video / Video generation
DiLA: Disentangled Latent Action World Models
Tianqiu Zhang, Muyang Lyu, Yufan Zhang, Fang Fang, Si 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 content pathway (disentanglement) to the structure-only latent action model (ablation) | DiLA · SSv2+RT-1 mixture | rollout LPIPS: 0.344 -> 0.263 | Improved |
| add content pathway (ablation) | DiLA · SSv2+RT-1 mixture | action cycle transfer LPIPS: 0.451 -> 0.343 | Improved |
| add content pathway (ablation) | DiLA · SSv2+RT-1 mixture | MSE after 10k iterations: 0.249 -> 0.216 | Improved |
| use temporal-difference bottleneck instead of NSVQ discrete bottleneck (ablation) | DiLA · SSv2+RT-1 mixture | rollout LPIPS: 0.334 -> 0.263 | Improved |
| use temporal-difference bottleneck instead of Gaussian (beta-VAE) bottleneck (ablation) | DiLA · SSv2+RT-1 mixture | rollout LPIPS: 0.346 -> 0.263 | Improved |
| use DiLA bottleneck instead of VQ or Gaussian for latent actions (ablation) | DiLA · OOD linear probing (Franka Kitchen) | probing MSE: 0.098 / 0.125 -> 0.073 | Improved |
| remove IDM and FDM (no latent action learning) (ablation) | DiLA · structure-only visualization and rebinding | disentanglement (qualitative) | Hurt |
| add content pathway to structure-only model (ablation) | DiLA · RT-1 rollout | SSIM: 0.647 -> 0.774 | Improved |
| compare DiLA with AdaWorld (external diffusion model) on SSIM | DiLA · SSv2 rollout | SSIM: 0.674 (AdaWorld) vs 0.660 (DiLA) | No clear effect |
| use DiLA instead of AdaWorld as dynamics model for MPC | visual planning with MPPI · VP2 benchmark aggregate | success rate (%): 63.50 -> 68.00 | 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.
- AdaWorld — AdaWorld: Learning Adaptable World Models with Latent Actions (ICML 2025) · better on 4 of 5 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
- Factored Latent Action World Models (ICML 2026 · similarity 0.91)
- Latent Action Learning Requires Supervision in the Presence of Distractors (ICML 2025 · similarity 0.88)
- Latent Action Pretraining from Videos (ICLR 2025 · similarity 0.88)
- Learning Visual Feature-Based World Models via Residual Latent Action (NeurIPS 2026 · similarity 0.88)
- DiT4DiT: Jointly Modeling Video Dynamics and Actions for Generalizable Robot Control (NeurIPS 2026 · similarity 0.87)
- SlowFast-VGen: Slow-Fast Learning for Action-Driven Long Video Generation (ICLR 2025 · similarity 0.87)
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.
