Light-X: Open Source 4D Video Camera & Relighting AI

Light-X (TQTQliu/Light-X) is an open source generative 4D video rendering framework developed by S-Lab (NTU), BAAI, Tsinghua University, and HUST (arXiv:2512.05115) that jointly controls 3D camera trajectories and scene illumination from a single monocular video. By decoupling dynamic 3D point cloud geometry from projected relighting cues, Light-X lets directors reshoot existing footage with gradual, bullet, direct, or Hitchcock style dolly-zoom camera moves while simultaneously relighting the scene from text prompts, background plates, reference images, or HDR environment maps.

Studio availability. Light-X is not available in AI FILMS Studio. This guide covers the open source research checkpoints and local workflow. In the Studio, use Runway Aleph, Kling O3 Video Edit, Wan 2.2 VACE, or Kling 2.6 Pro Motion Control on the Video to Video page or in Studio Nodes.

How Light-X Disentangles 4D Geometry and Illumination

Most video relighting models require a locked or original camera path, while camera trajectory models (such as TrajectoryCrafter) bake the source lighting into the point cloud and cannot alter shadows or specular highlights when the camera orbits.

Light-X solves both problems in a single generative pass through two technical contributions.

1. Disentangled geometry and lighting projection. Geometry and subject motion are lifted from the monocular video into dynamic 3D point clouds and projected along the user's target camera trajectory (--camera traj or --camera target). Simultaneously, a relit anchor frame is projected into the exact same 3D geometry so the Diffusion Transformer receives spatially aligned, decoupled geometry and illumination cues.
2. The Light-Syn training dataset (tqliu/Light-Syn). Because studios rarely capture synchronized multiple view and multiple illumination video pairs outside of a Light Stage, the authors built a degradation and inverse mapping pipeline that synthesizes paired training clips from monocular footage.

Four Camera Modes & Four Relighting Sources

Light-X exposes cinematic camera choreography and lighting controls directly through CLI flags in inference.py.

• Camera motion styles (--mode).
  • gradual: Smooth, continuous viewpoint transition for natural dolly or crane moves.
  • bullet: Fast orbit and parallax shift around the subject for slow motion or frozen moment shots.
  • direct: Rapid transition from the starting camera pose to a target <theta phi r x y> pose.
  • dolly-zoom: Simultaneous camera translation and radius adjustment (Hitchcock vertigo zoom) that keeps the foreground subject size constant while compressing or expanding the background.
• Illumination sources (--relit_cond_type).
  • ic: Text prompted or background plate conditioned relighting (tqliu/Light-X).
  • ref: Reference image lighting transfer (tqliu/Light-X-Uni).
  • hdr: 360 degree HDR environment map illumination (tqliu/Light-X-Uni).
  • bg: Composite background conditioning with automatic foreground masking (--mask).

Checkpoints, Dataset & Hardware Matrix

Resource	Hugging Face / GitHub ID	Supported Conditioning	Hardware & VRAM	License
Light-X Base Model	tqliu/Light-X	Text prompt + background image relighting + 3D camera control	Single 24 to 48 GB VRAM NVIDIA GPU (RTX 4090 / A6000 / A100)	Open Research Release
Light-X-Uni Model	tqliu/Light-X-Uni	HDR environment map + reference image + text + background	Single 24 to 48 GB VRAM NVIDIA GPU	Open Research Release
Light-Syn Dataset	tqliu/Light-Syn	Paired multiple view and multiple illumination synthetic video dataset	Training & evaluation dataset	Open Research Dataset

Local Installation & CLI Usage

Clone the repository and create the Python 3.10 Conda environment. Pretrained weights from tqliu/Light-X and tqliu/Light-X-Uni download automatically on first run.

git clone https://github.com/TQTQliu/Light-X.git
cd Light-X
conda create -n lightx python=3.10 -y
conda activate lightx
pip install -r requirements.txt

Run joint camera trajectory rendering (--recam_vd) and HDR or text relighting (--relit_vd).

python inference.py \
  --video_path test/videos/1.mp4 \
  --stride 1 \
  --out_dir outputs/dolly_zoom_relight \
  --camera traj \
  --mode dolly-zoom \
  --mask \
  --traj_txt test/trajs/loop1.txt \
  --relit_txt test/Prompts/relight_prompts.txt \
  --relit_cond_type ic \
  --relit_vd \
  --recam_vd

For actor relighting where preserving exact expressions and lip sync is the top priority, compare Light-X with Netflix and Eyeline Labs' ID-V2V and NVIDIA UniRelight. For dedicated camera trajectory control, see CameraNoise, TriMotion, and FlashMotion or InfCam precise camera control.

Camera Control and Video Relighting in AI FILMS Studio

Light-X runs locally on CUDA GPUs. To change camera framing and transform scene lighting in the browser without building local point clouds, open the video workspace to run Runway Aleph, Kling O3 Video Edit, Kling 2.6 Pro Motion Control, and Wan 2.2 VACE, or connect camera and relighting passes in the Nodes Graph Editor.

---

Sources

• TQTQliu/Light-X GitHub Repository
• Light-X Official Project Page
• Light-X: Generative 4D Video Rendering with Camera and Illumination Control (arXiv:2512.05115)
• tqliu/Light-X on Hugging Face
• tqliu/Light-X-Uni on Hugging Face