WORL-Lingbot Video Generation Model: File List

Sample outputs are available on the model card.
At a Glance
| Item | Value |
|---|---|
| Repository | Efficient-Large-Model/WORL-Lingbot |
| Publisher guide | Efficient Large Model (NVIDIA and MIT): models and licenses |
| Published | 2026-10-20 |
| License | apache-2.0 |
| Formats | safetensors |
| Source type | Primary source (the publisher itself) |
Values determined by this site’s code when the information was collected. Dates are JST.
Overview
Efficient-Large-Model has released “WORL-Lingbot", a LoRA adapter designed for video world models.
This model is a camera-control LoRA adapter trained by applying the “WORL" reinforcement learning method to the base model, “LingBot-World-Fast". By taking an initial image (RGB image), text prompt, camera trajectory for each frame, and intrinsic parameter matrices as inputs, it outputs continued video generation results following the specified camera work.
Specifications
Specifications listed in the model card and accompanying metadata are as follows:
- Architecture: Camera-conditioned
WanModelFasttransformer - Base model:
robbyant/lingbot-world-fast(revisionb7a1db2da9bd3ba675ee01c228f41db0309b981e, subfolderlingbot_world_fast) - Distribution format: BF16 precision PEFT LoRA adapter (safetensors format)
- LoRA parameters: rank 128 / alpha 128
- Output specs: 81 frames, resolution 832 × 464 pixels, frame rate 16 FPS (specified per the original release records)
- Evaluation stride: A stride of 80 frames between 81-frame clips, sharing 1 boundary frame between adjacent clips
- Input data formats:
- First frame: RGB image – Prompt: Text describing the scene and intended movement – Camera trajectory: Camera-to-world transformation matrix per frame
(F, 4, 4)– Camera intrinsics: Intrinsic parameter matrix per frame(F, 3, 3) - Required components: Does not operate standalone; separately requires the base transformer, text encoder, tokenizer, VAE, scheduler, and LingBot runtime
- License conditions: Released under the Apache-2.0 license
Performance and Quality
While the model card does not include a numerical comparison table of benchmark scores, training methodology, quality characteristics, and the published performance of the base model have been reported.
In WORL, the training methodology used for this model, policy optimization is performed through steps of autoregressive rollouts, reward evaluation, candidate selection, and policy updates. The reference configuration explains that training was conducted by combining HPSv3, NVRM, and VGGT-Omega PI Camera reward functions. As a training dataset, the “OpenScene/revisit training bundle" (qyoo/lingbot-top2400-training-data) consisting of 2,218 records containing prompts, initial frames, and camera conditioning arrays was prepared, and gpt-image was stated to be used for initial scene images.
According to information published by the Robbyant team regarding the base model “LingBot-World-Fast" (LingBot-World series), it is reported to maintain high fidelity and robust dynamics across diverse environments, ranging from realistic representations to scientific contexts and cartoon styles. It is also described as featuring long-term memory capabilities for minute-long predictive generation while maintaining context consistency, and real-time interactivity achieving sub-second latency at 16 FPS generation.
On the other hand, the publishers have clearly noted limitations in use. It is reported that long-duration rollouts may lead to drift (deviations or disruptions) in scene identity, spatial geometry, camera motion, and physical consistency. Furthermore, cautions are raised that generated synthetic video should not be treated as verified physics simulation or autonomous control signals, and that usage requires verification due to potential biases originating from the base model and training data. Note that no new GPU benchmarks were conducted for this release.
Strengths and Use Cases
This model is designed for primary use cases including camera trajectory-conditioned video generation technology and reinforcement learning (RL) research for video world models. Beyond mere text-to-video generation, a major feature is its ability to perform continued video generation with precisely controlled viewpoint movement by specifying camera trajectories (camera-to-world matrices) and intrinsic parameters per frame, in addition to a reference single image (first frame) and text scene instructions.
This LoRA adapter is based on the fast-inference transformer “LingBot-World-Fast" and aims to enhance camera-following and world simulation characteristics through additional training using rollouts and reward feedback via the WORL method. Strengths highlighted for the base model “LingBot-World" series include high expressive power capable of handling diverse spaces such as realistic live-action video expressions, scientific simulation environments, and cartoon-style animations. The publishers envision these technologies contributing to research and development in practical fields such as content creation, game development, and robot learning.
Meanwhile, the model card also notes cautions for use. Repeated long-term autoregressive rollouts tend to cause drift (disruptions or deviations) in object identity, spatial geometry, camera movement, and physical behavior consistency within the scene. Therefore, practical deployment as a verified rigorous physics simulator or autonomous control signal for autonomous driving or robotics is not intended, and users are required to verify whether generated results are suitable for their intended use and to consider the rights of input media.
Distributed Files
Weight files published in Efficient-Large-Model/WORL-Lingbot, listed by this site from the Hugging Face API. Sizes are the actual file sizes.
| File | Size |
|---|---|
adapter_model.safetensors |
1.23GB |
How to Get It
WORL-Lingbot is available on the Hugging Face repository “Efficient-Large-Model/WORL-Lingbot". As it is not a gated model, it can be directly downloaded and used without prior application or consent procedures. The Apache-2.0 license applies.
It is distributed as a BF16 precision PEFT LoRA adapter. The repository contains the adapter weights adapter_model.safetensors, the PEFT configuration file adapter_config.json, and RELEASE_MANIFEST.json recording revision information and SHA-256 hashes. According to the publisher’s explanation, the adapter weights and configuration contents are identical to a specific revision (d3e449c675f5) of qyoo/lingbot-worl, and file integrity has been verified against that snapshot.
The weight files can be downloaded using the official Hugging Face CLI tool. You can fetch them to a local directory by running the following command in your terminal:
hf download Efficient-Large-Model/WORL-Lingbot --local-dir artifacts/WORL-Lingbot
to actually execute video generation, you will separately need a compatible LingBot runtime environment along with the base transformer, text encoder, tokenizer, VAE, and scheduler, rather than just this adapter alone. The base model codebase is based on Wan2.2, and is operated in a Python environment after setting up dependent libraries such as PyTorch (version 2.4.0 or higher) and flash-attn.
When loading the model from Python code, use the PEFT library to attach the adapter to a fixed, specific revision of the LingBot-World-Fast transformer. The loading example presented in the model card is as follows:
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()
After loading the base transformer and LoRA adapter in this manner, you can integrate them with LingBot pipeline components to build a video generation pipeline specifying camera conditioning.
Can You Run It Locally?
The publisher distributes this model as safetensors.
License — apache-2.0 (Commercial use allowed): Permits commercial use, modification and redistribution. Redistribution requires including the license and stating changes; includes a patent grant.
Compiled by this site’s code from the published formats and the license field. License summaries are not legal advice — check the publisher’s original terms before relying on them.
Related Articles
- WORL-Helios Video Generation Model: File List
- LongLive-Plug-Wan2.1-T2V-14B-few-step: Our Generated Video, File List
- LongLive-Plug-Wan2.1-T2V-14B-cfg: Our Generated Video, File List
- LongLive-Plug-Wan2.2-TI2V-5B-cfg: Our Generated Video, File List
What to Read Next
- Formats this model is available in → Safetensors format guide and models
- Learn about the publisher → Efficient Large Model (NVIDIA and MIT): models, licenses and articles
- Other models for the same task → Other video generation models

