Elastic Forcing
Elastic Forcing post-trains autoregressive video generators by matching generated and reference videos in frozen representation spaces. It combines hybrid Nyström–Monte Carlo distribution estimation with gradient replay.
Model checkpoints · 1.3B guide · 14B guide
Models
| Backbone | Frozen feature encoders | Training GPUs | Training updates | Inference |
|---|---|---|---|---|
| Wan2.1-T2V-1.3B | DINOv3 + V-JEPA 2 + VideoMAEv2 | 4 | 150 | Four causal denoising steps per chunk |
| Wan2.1-T2V-14B | DINOv3 + V-JEPA 2 + VideoMAEv2 | 8 | 80 | Four causal denoising steps and terminal cleanup |
Both models generate 81-frame videos at 832 × 480 resolution and 16 FPS. Training uses a distribution batch of 256 generated videos, a differentiation batch of 128 generated videos, and a reference minibatch of 128 videos.
Getting started
git clone https://huggingface.co/Elastic-forcing/Elastic-forcing-code
cd Elastic-forcing-code
# Select the model size and follow its guide.
cd 1.3B
# Or: cd 14B
Each model directory includes its dependencies, configuration, asset guide, training launcher, and inference launcher. Download the matching generator checkpoint from the checkpoint repository and prepare the Wan2.1 base-model assets locally. Training additionally requires an ODE initialization, frozen feature encoders, and reference features.
Elastic-forcing-code/
├── 1.3B/
│ ├── code/
│ ├── configs/
│ ├── scripts/
│ └── README.md
├── 14B/
│ ├── code/
│ ├── configs/
│ ├── scripts/
│ └── README.md
└── docs/
The model guides describe dependencies, required assets, training settings, and inference commands.
Evaluation
| Model | VBench Total ↑ | Quality ↑ | Semantic ↑ |
|---|---|---|---|
| Elastic Forcing 1.3B | 84.25 | 85.06 | 80.99 |
Per-dimension scores are provided in the reference results. Evaluate generated videos using the official VBench implementation.
Acknowledgments and license
This implementation builds on Self-Forcing, Wan2.1, DINOv3, V-JEPA 2, and VideoMAEv2.
Source-code terms are provided in LICENSE. Upstream source notices are retained; model assets and datasets follow their respective upstream terms.