---
license: apache-2.0
base_model: robbyant/lingbot-world-fast
library_name: peft
pipeline_tag: image-to-video
inference: false
tags:
  - worl
  - lingbot
  - world-model
  - video-generation
  - camera-control
  - reinforcement-learning
  - lora
  - safetensors
---

# WORL-Lingbot

**WORL-Lingbot** is a camera-conditioned video world-model LoRA adapter trained with
WORL on top of [LingBot-World-Fast](https://huggingface.co/robbyant/lingbot-world-fast).
It generates video continuations conditioned on an initial image, text and
camera trajectories.

<img src="assets/worl-teaser.webp" alt="WORL long-horizon video generation examples" width="100%">

Related releases: [WORL-NVRM](https://huggingface.co/Efficient-Large-Model/WORL-NVRM)
and [WORL-Helios](https://huggingface.co/Efficient-Large-Model/WORL-Helios).

## Model details

| Field | Description |
| --- | --- |
| Base model | `robbyant/lingbot-world-fast` |
| Base revision | `b7a1db2da9bd3ba675ee01c228f41db0309b981e` |
| Base subfolder | `lingbot_world_fast` |
| Architecture | Camera-conditioned `WanModelFast` transformer |
| Release format | BF16 PEFT LoRA adapter |
| LoRA rank / alpha | 128 / 128 |
| Output geometry documented by the source release | 81 frames, 832 × 464 pixels, 16 FPS |

Load this adapter on top of the pinned LingBot-World-Fast base transformer.
The compatible base transformer, text encoder, tokenizer, VAE, scheduler and
LingBot runtime are required separately.

## Inputs and outputs

| Input | Format |
| --- | --- |
| First frame | RGB image |
| Prompt | Text describing the scene and intended motion |
| Camera trajectory | Per-frame camera-to-world matrices, `(F, 4, 4)` |
| Camera intrinsics | Per-frame intrinsic matrices, `(F, 3, 3)` |

The output is a camera-conditioned video continuation. The source evaluation
profile uses an 80-frame stride between 81-frame clips, sharing one boundary
frame between adjacent clips.

## Training data and method

WORL trains policies using autoregressive rollouts, reward evaluation,
candidate selection and policy updates. The project's LingBot reference
configuration combines HPSv3, NVRM and VGGT-Omega PI Camera rewards.

The team supplied [WORL-Lingbot-Training-Data](https://huggingface.co/datasets/Efficient-Large-Model/WORL-Lingbot-Training-Data), an OpenScene/revisit training bundle
containing 2,218 records with prompts, initial frames and camera-conditioning
arrays. Initial scene imagery was generated using gpt-image. This describes
the supplied project dataset and reference configuration; the exported model
does not record an exact per-run dataset/configuration binding. The training
dataset is distributed separately under CC BY 4.0.

## Download and load

```bash
hf download Efficient-Large-Model/WORL-Lingbot --local-dir artifacts/WORL-Lingbot
```

With a compatible LingBot runtime installed, attach the adapter to the pinned
base transformer:

```python
import torch
from peft import PeftModel
from wan.modules.model_fast import WanModelFast

base = WanModelFast.from_pretrained(
    "robbyant/lingbot-world-fast",
    revision="b7a1db2da9bd3ba675ee01c228f41db0309b981e",
    subfolder="lingbot_world_fast",
    torch_dtype=torch.bfloat16,
    control_type="cam",
)
model = PeftModel.from_pretrained(
    base,
    "Efficient-Large-Model/WORL-Lingbot",
    is_trainable=False,
).eval()
```

This example loads the base transformer and adapter. Integrate them with the remaining compatible
pipeline components for complete video generation.

## Files and provenance

| Path | Contents |
| --- | --- |
| `adapter_model.safetensors` | Approximately 1.23 GB BF16 adapter |
| `adapter_config.json` | PEFT configuration |
| `RELEASE_MANIFEST.json` | Source revision and SHA-256 inventory |

Adapter weights and configuration are unchanged from
[qyoo/lingbot-worl at d3e449c675f5](https://huggingface.co/qyoo/lingbot-worl/tree/d3e449c675f5a8bfb3afd29e30f1675e4445b9c3).
Documentation and release manifests were updated for this repository's
adapter-only release. The original export record remains available at the
pinned source revision. File integrity is checked against that source snapshot;
no new GPU benchmark was performed for this publication.

## Intended use and limitations

This model supports research on camera-conditioned video generation and video
world-model reinforcement learning. Long rollouts can drift in scene identity,
geometry, camera motion and physical consistency. Synthetic outputs are not
validated physical simulations or autonomous-control signals. Base-model and
training-data biases can affect generations; review outputs for the intended
application and respect rights in input media.

## License

This release retains the source model card's
[Apache License 2.0](LICENSE.txt). The base model and separately obtained runtime
components retain their upstream terms and notices. See [NOTICE.txt](NOTICE.txt)
for source attribution.