---
pipeline_tag: text-to-video
tags:
- text-to-video
- pytorch
- safetensors
---
# Elastic Forcing Checkpoints

Generator checkpoints for autoregressive text-to-video generation. Use them with the [Elastic Forcing code](https://huggingface.co/Elastic-forcing/Elastic-forcing-code).

| Model | Training updates | Format | Precision |
|---|---:|---|---|
| 1.3B | 150 | `checkpoints/1.3B/model.pt` | FP32 |
| 14B | 80 | Eight safetensors shards and an index under `checkpoints/14B/` | FP32 |

Inference requires the matching Wan2.1 base-model assets, T5 text encoder, tokenizer, and VAE. The frozen feature encoders are required for training. Use the `generator` weights without EMA.

## Download

```python
from huggingface_hub import snapshot_download

folder = snapshot_download(
    "Elastic-forcing/Elastic-forcing-checkpoint",
    allow_patterns=["checkpoints/1.3B/*"],  # Select checkpoints/14B/* for the 14B model.
)
print(folder)
```

## Inference

Follow the [1.3B guide](https://huggingface.co/Elastic-forcing/Elastic-forcing-code/blob/main/1.3B/README.md) or [14B guide](https://huggingface.co/Elastic-forcing/Elastic-forcing-code/blob/main/14B/README.md) to prepare the environment and local base-model assets.

For 1.3B, set `CHECKPOINT` to the downloaded `model.pt`. For 14B, combine the safetensors shards into the generator checkpoint expected by the launcher:

```python
import json
from pathlib import Path
import torch
from safetensors.torch import load_file

folder = Path("/path/to/checkpoints/14B")
index = json.loads((folder / "model.safetensors.index.json").read_text())
state = {}
for shard in sorted(set(index["weight_map"].values())):
    state.update(load_file(str(folder / shard), device="cpu"))
torch.save({"generator": state}, "/path/to/local/model.pt")
```

The conversion preserves FP32 tensor values and needs sufficient host memory and disk space for the approximately 57 GB model and its output file. Set `CHECKPOINT` to the resulting file and run the corresponding inference launcher.

Source-code license notices are provided in the code repository. Base-model assets and datasets follow their respective upstream terms.