LingBot-World#

Project page

Official code

Introduced by Robbyant, LingBot-World is a camera-controllable image-to-video (I2V) world model with streaming inference and context-parallel runtime support. This page covers both the original LingBot-World v1 and the newer 14B causal-fast LingBot-World v2 checkpoints.

Teaser video source: LingBot-World project page.

Requirements#

  • Minimum VRAM: ~120 GB.

  • PyTorch: >= 2.9.

Installation#

# from the repo root
uv sync --project integrations_v2/lingbot

Running the method#

The Lingbot package binds its model pipeline directly to the reusable Cam2V v2 application:

uv sync --package flashdreams-lingbot --inexact
uv run --no-sync flashdreams-run-v2 cam2v-lingbot \
    --mode webrtc --host 0.0.0.0 --port 8089 -- --example-data

Application arguments follow --. Run flashdreams-run-v2 cam2v-lingbot -- --help for custom first-frame, intrinsics, prompt, and motion-normalizer inputs.

Sample data is downloaded from the LingBot-World v2 repository. Valid --example-idx values are 0, 1, 2, 5. Note the single GPU command might run out of memory for large --total-blocks values.

The package exposes the following pipeline configurations for programmatic use:

Method

Description

lingbot-world-fast

Official camera-control I2V (Wan VAE decoder, full KV-cache).

lingbot-world-fast-taehv-window15-sink3

Efficient streaming configuration: TAEHV decoder, window_size_t=15 + sink_size_t=3 streaming KV-cache.

lingbot-world-v2-14b-causal-fast

LingBot-World V2 14B causal-fast on the shared LingBot pipeline (Wan VAE decoder, 4-step). See LingBot-World V2.

lingbot-world-v2-14b-causal-fast-taehv-window15-sink3

LingBot-World V2 14B causal-fast with the TAEHV decoder, window_size_t=15 + sink_size_t=3 streaming KV-cache.

LingBot-World V2#

LingBot-World V2 is the newer 14B causal-fast checkpoint from Robbyant. It uses the same architecture and pipeline code as v1; only the checkpoint configuration changes. See the canonical repository at Robbyant/lingbot-world-v2.

Two V2 pipeline configs are exported:

Method

Description

lingbot-world-v2-14b-causal-fast

LingBot-World V2 14B causal-fast on the shared LingBot pipeline (Wan VAE decoder, full KV-cache).

lingbot-world-v2-14b-causal-fast-taehv-window15-sink3

V2 checkpoint with the efficient streaming preset: TAEHV decoder, window_size_t=15 + sink_size_t=3 streaming KV-cache.

The V2 checkpoint (~70 GB) is pulled from huggingface.co/robbyant/lingbot-world-v2-14b-causal-fast on first run. Export HF_TOKEN first.

What to expect#

  • Example data: --example-data True downloads image.jpg, intrinsics.npy, poses.npy, prompt.txt from the canonical examples folder into assets/example_data/lingbot_world/<NN>/ (<NN> matches --example-idx). Cached after first run; no credentials needed.

  • Model checkpoint: ~70 GB pulled from huggingface.co/robbyant/lingbot-world-fast on first run, cached under $HF_HOME. Export HF_TOKEN first.

  • Disk: keep ~200 GB free for the model + HF cache. Hosts under ~100 GB have been seen to run out mid-load.

  • First launch: a few minutes (download + Triton autotuning + CUDA-graph warmup). Subsequent launches reuse the caches.

  • Outputs: select MP4, WebRTC, or native-window presentation with the flashdreams-run-v2 runtime arguments. The Cam2V defaults are 16 FPS and 464×832.

See Inference pipeline overview for what one autoregressive chunk does end-to-end.

Some generated samples from the above commands:

example_idx: 01
example_idx: 02

Launch the interactive server#

Run the same Cam2V application in WebRTC mode:

uv run --no-sync flashdreams-run-v2 cam2v-lingbot \
    --mode webrtc --host 0.0.0.0 --port 8089 -- --example-data

When successfully connected, the browser-based UI looks like this:

Profiling benchmark#

Here is the profiling benchmark on total DiT runtime for FlashDreams LingBot-World compared to the official LingBot-World implementation and LightX2V under matched settings.

This chart shows total DiT runtime (4 diffusion steps) in milliseconds at the 6th autoregressive rollout on 4x GPUs. For an apples-to-apples comparison, all implementations are forced to use cuDNN attention backend under matched runtime settings, and all runs use Ulysses sequence parallelism for multi-GPU inference. For the official LingBot-World implementation, see this instruction. For the LightX2V baseline, see this instruction.

Citation#

If you use LingBot-World, please cite the original work:

@article{lingbot-world,
      title={Advancing Open-source World Models},
      author={Robbyant Team and Zelin Gao and Qiuyu Wang and Yanhong Zeng and Jiapeng Zhu and Ka Leong Cheng and Yixuan Li and Hanlin Wang and Yinghao Xu and Shuailei Ma and Yihang Chen and Jie Liu and Yansong Cheng and Yao Yao and Jiayi Zhu and Yihao Meng and Kecheng Zheng and Qingyan Bai and Jingye Chen and Zehong Shen and Yue Yu and Xing Zhu and Yujun Shen and Hao Ouyang},
      journal={arXiv preprint arXiv:2601.20540},
      year={2026}
}