LingBot-World camera pose control, and the notice that sends you to v2

The README opens by telling you to leave

The first thing this repository does is redirect you. A callout at the top of the README announces that LingBot-World-Infinity is available, states that this repository is no longer actively maintained, and points at https://github.com/Robbyant/lingbot-world-v2 as the place where all future updates, new models, and new features will be released.

Nothing is hidden from you. The archive flag is still off and the last recorded push is 2026-07-09, so the tree is reachable, but the code, the model table, and the news list describe a stopped snapshot. The repository also has no GitHub releases, which means there is no tag to pin, no release notes to read, and no versioned artifact to download. Every reproducible fact on this page traces back to the news entries and the reference inference scripts, whose dated entries stop at Apr 10, 2026.

Control arrives as intrinsics.npy and poses.npy, in OpenCV coordinates

The control interface is camera geometry rather than text. Before inference runs, the README asks for three things: an input image, a text prompt, and control signals that are marked optional and can be generated from a video with ViPE (https://github.com/nv-tlabs/vipe).

Those signals are two NumPy files with fixed shapes:

- `intrinsics.npy`: shape [num_frames, 4], the four values are [fx, fy, cx, cy] - `poses.npy`: shape [num_frames, 4, 4], each [4, 4] is a transformation matrix in OpenCV coordinates

The coordinate choice matters. Poses are expressed in OpenCV convention, so a pose stream produced by a renderer that uses the opposite handedness needs conversion before it is placed in the directory passed to `--action_path`. The repository ships examples/00 through examples/05 as ready made signal folders, and examples/ also contains persistent_inference.py.

Eight ranks, 161 frames, and one torchrun line

The 480P reference invocation for LingBot-World-Base (Cam) is a single distributed command:

torchrun --nproc_per_node=8 generate.py --task i2v-A14B --size 480*832 --ckpt_dir lingbot-world-base-cam --image examples/00/image.jpg --action_path examples/00 --dit_fsdp --t5_fsdp --ulysses_size 8 --frame_num 161 --prompt "The video presents a soaring journey through a fantasy jungle. The wind whips past the rider's blue hands gripping the reins, causing the leather straps to vibrate. The ancient gothic castle approaches steadily, its stone details becoming clearer against the backdrop of floating islands and distant waterfalls."

Read the flags as one hardware decision rather than ten. `--nproc_per_node=8` asks for eight processes, `--dit_fsdp` and `--t5_fsdp` shard the diffusion transformer and the text encoder across those ranks, and `--ulysses_size 8` sets sequence parallelism to the same width. `--frame_num 161` fixes the output length, `--action_path examples/00` supplies the intrinsics and poses, and `--task i2v-A14B` is passed through from the Wan2.2 codebase this project is built on. The 720P line keeps every flag and changes `--size` to 720*1280. Other entry points sit in the root: generate_fast.py, run_fast.sh, run_act2cam.sh, and run_act2cam_string.sh.

Two dependency lists that disagree about transformers

The Quick Start leans on another project for setup, then gets very short. It states the codebase is built upon Wan2.2 and sends you to that project's documentation for installation instructions, so the sequence here is only:

git clone https://github.com/robbyant/lingbot-world.git
cd lingbot-world

# Ensure torch >= 2.4.0
pip install -r requirements.txt

pip install flash-attn --no-build-isolation

The last command is worth noting because flash_attn already appears inside both dependency lists, so the separate install is not redundant by accident. The two lists also differ. requirements.txt pins `transformers>=4.49.0,<=4.51.3` and adds torchaudio plus `imageio[ffmpeg]`, while pyproject.toml declares `transformers>=4.49.0` with no ceiling and plain imageio. Both pin opencv-python>=4.9.0.80, diffusers>=0.31.0, accelerate>=1.1.1, and numpy>=1.23.5,<2. pyproject.toml requires Python >=3.10,<4.0. If you install from the package metadata instead of requirements.txt you get a different transformers range.

One download link covers two different models

Three checkpoints are offered, and the table that lists them points two rows at the same place.

| Model | Control Signals | Resolution | | :--- | :--- | :--- | | LingBot-World-Base (Cam) | Camera Poses | 480P & 720P | | LingBot-World-Base (Act) | Actions | 480P & 720P | | LingBot-World-Fast | Camera Poses | 480P & 720P |

LingBot-World-Base (Cam) is listed on both HuggingFace at robbyant/lingbot-world-base-cam and on ModelScope, and LingBot-World-Fast points at robbyant/lingbot-world-fast, but the Base (Act) row, the one that takes actions instead of camera poses, carries the same HuggingFace path as the Cam row. So the signals you feed and the weights you fetch are decided by two different rows of the same table, and the Act row gives you nowhere new to download from.

pip install "huggingface_hub[cli]"
huggingface-cli download robbyant/lingbot-world-base-cam --local-dir ./lingbot-world-base-cam

pip install modelscope
modelscope download robbyant/lingbot-world-base-cam --local_dir ./lingbot-world-base-cam

The two commands are near twins, differing only in the directory flag, `--local-dir` for the HuggingFace CLI and `--local_dir` for ModelScope.

The claims are three numbers wide, and the hardware story is empty

The feature list is short and quantitative: high fidelity across realism, scientific contexts, and cartoon styles, a minute-level horizon with contextual consistency over time, and latency under 1 second when producing 16 frames per second.

The reference inference commands name no GPU model, VRAM figure, guidance scale or sampler step count.

The consequence for a reader is concrete. The eight rank FSDP command tells you the parallelism the authors used, not the hardware it needs, so any capacity planning before your first run is your own estimate rather than a number the reference commands supply.

The paper, the license, and the dates in the tree

What you can audit lives in the repository rather than in a release feed. Root entries include LingBot_World_paper.pdf, LICENSE.txt, pyproject.toml, requirements.txt, generate.py, generate_fast.py, run_fast.sh, run_act2cam.sh, run_act2cam_string.sh, plus assets/, examples/, and the wan/ package. The package metadata declares the name lingbot-world at version 1.0.0 under the Apache-2.0 license file, publishes packages = ["wan"], and lists dev extras with pytest, black, flake8, isort, mypy, and huggingface-hub[cli]. Formatting is pinned too: black uses line-length 88, isort uses profile black, and mypy is set to strict.

The news list dates the work precisely: technical report, code, and models on Jan 29, 2026, Base (Act) weights on Mar 2, Fast weights on Apr 2, Fast inference scripts on Apr 7, and user friendly Base (Act) scripts on Apr 10. Reactor provides an online demo at https://www.reactor.inc/. The homepage is https://technology.robbyant.com/lingbot-world, the report is at arxiv.org/abs/2601.20540, and a HuggingFace collection groups the weights. Any documentation fix you need now belongs in the v2 repository, not here.

Editorial conclusion

Read this repository as a frozen snapshot rather than an active project. Its useful facts are concrete: camera intrinsics and poses in OpenCV coordinates, one torchrun line with FSDP and Ulysses sharding, 480P and 720P size strings, and a dated news list that stops in April 2026. The reference commands state no memory footprint, GPU model or step count, and the transformers ceiling differs between requirements.txt and pyproject.toml. Check lingbot-world-v2 before you plan any work against this code.

Frequently asked questions

Official sources

1. Issues
2. License: Apache-2.0
3. Project website
4. README
5. Robbyant/lingbot-world on GitHub