nexuslrf/luxdit_video

LuxDiT

This is the luxdit_video checkpoint for LuxDiT: Lighting Estimation with Video Diffusion Transformer. It is finetuned on video data and includes a LoRA adapter for real scenes.

Model description

LuxDiT is a generative lighting estimation model that predicts high-quality HDR environment maps from visual input. It produces accurate lighting while preserving scene semantics, enabling realistic virtual object insertion under diverse lighting conditions. This model is ready for non-commercial use.

• Checkpoint: Video model (video-finetuned)
• LoRA: Included for real-scene generalization
• Paper: LuxDiT: Lighting Estimation with Video Diffusion Transformer
• Project page: https://research.nvidia.com/labs/toronto-ai/LuxDiT/

Use case: LuxDiT supports studies and prototyping in video lighting estimation. This release is an open-source implementation of our research paper, intended for AI research, development, and benchmarking for lighting estimation research.

Model architecture

• Architecture type: Transformer (based on CogVideoX)
• Parameters: 5B
• Input: RGB video frames; shape [batch_size, num_frames, height, width, 3]; recommended resolution 480×720
• Output: RGB video frames (dual tonemapped LDR and log); output resolution 256×512; use the HDR merger to obtain .exr HDR envmaps

Software and hardware

• Runtime: Python and PyTorch
• Supported hardware: NVIDIA Ampere (e.g. A100 GPUs)
• Operating system: Linux

How to use

Download from Hugging Face

From the LuxDiT repository root:

python download_weights.py --repo_id <HF_ORG>/luxdit_video

This saves the checkpoint to checkpoints/luxdit_video by default. Use --local_dir to override.

Inference: synthetic videos (in-domain)

Requires --data_type video:

DIT_PATH=checkpoints/luxdit_video
INPUT_DIR=examples/input_demo/synthetic_videos
OUTPUT_DIR=test_output/synthetic_videos

python inference_luxdit.py \
    --config configs/luxdit_base.yaml \
    --transformer_path $DIT_PATH \
    --input_dir $INPUT_DIR \
    --output_dir $OUTPUT_DIR \
    --resolution 480 720 \
    --guidance_scale 2.5 \
    --num_inference_steps 40 \
    --seed 33 \
    --data_type video

python hdr_merger.py \
    --input_dir $OUTPUT_DIR/ldr_log \
    --output_dir $OUTPUT_DIR/hdr

Inference: real scenes from video (with LoRA)

Use the LoRA adapter in this checkpoint for better generalization to real video:

DIT_PATH=checkpoints/luxdit_video
LORA_PATH=checkpoints/luxdit_video/lora
INPUT_DIR=examples/input_demo/scene_videos
OUTPUT_DIR=test_output/scene_videos

python inference_luxdit.py \
    --config configs/luxdit_base.yaml \
    --transformer_path $DIT_PATH \
    --lora_dir $LORA_PATH \
    --lora_scale 0.8 \
    --input_dir $INPUT_DIR \
    --output_dir $OUTPUT_DIR \
    --resolution 480 720 \
    --guidance_scale 2.5 \
    --num_inference_steps 40 \
    --seed 33 \
    --data_type video

python hdr_merger.py \
    --input_dir $OUTPUT_DIR/ldr_log \
    --output_dir $OUTPUT_DIR/hdr

Adjust lora_scale (e.g. 0.0–1.0) to control how much the input scene is merged into the estimated envmap.

Inference: object video scan with camera poses

For multi-view object captures, you can optionally provide camera poses (--camera_pose_file, OpenCV format) to align the estimated envmap to a canonical layout:

DIT_PATH=checkpoints/luxdit_video
LORA_PATH=checkpoints/luxdit_video/lora
INPUT_DIR=examples/input_demo/object_scans/antman
OUTPUT_DIR=test_output/object_scans/antman
CAM_FILE=examples/input_demo/object_scans/antman/antman.camera.json

python inference_luxdit.py \
    --config configs/luxdit_base.yaml \
    --transformer_path $DIT_PATH \
    --lora_dir $LORA_PATH \
    --lora_scale 0.0 \
    --input_dir $INPUT_DIR \
    --output_dir $OUTPUT_DIR \
    --resolution 512 512 \
    --guidance_scale 2.5 \
    --num_inference_steps 40 \
    --seed 33 \
    --data_type video \
    --camera_pose_file $CAM_FILE

python hdr_merger.py \
    --input_dir $OUTPUT_DIR/ldr_log \
    --output_dir $OUTPUT_DIR/hdr

Related checkpoints

Checkpoint	Description
luxdit_base	Image-finetuned, with LoRA for real scenes
luxdit_video (this model)	Video-finetuned, with LoRA for real scenes

For single-image inputs, use luxdit_base instead.

Training data

This checkpoint is video-finetuned on the SyntheticScenes dataset:

• Modality: Video
• Scale: ~108,000 rendered videos; each video has 57 frames at 704×1280 resolution
• Collection: Synthetic data generated with an OptiX-based physically based path tracer
• Labels: Produced by the renderer (no manual labeling)
• Per sample: Input RGB (LDR) video and HDR environment lighting

Testing and evaluation use held-out 10% splits of the same dataset.

Output format

The model outputs dual tonemapped environment maps (LDR and log); use the HDR merger to get .exr HDR envmaps. By default, the camera pose of the input image defines the world frame. See the main README for the exact layout of LDR/log vs merged HDR.

Ethical considerations

NVIDIA believes Trustworthy AI is a shared responsibility. When using this model in accordance with the terms of service, ensure it meets requirements for your use case and addresses potential misuse. You are responsible for having proper rights and permissions for all input image and video content; if content includes people, personal health information, or intellectual property, generated outputs will not blur or preserve proportions of subjects. Users are responsible for model inputs and outputs and for implementing appropriate guardrails and safety mechanisms before deployment. To report model quality, risk, security vulnerabilities, or other concerns, see NVIDIA AI Concerns.

License

NVIDIA OneWay Noncommercial License. See the LICENSE in the LuxDiT repository.

Citation

@article{liang2025luxdit,
  title={Luxdit: Lighting estimation with video diffusion transformer},
  author={Liang, Ruofan and He, Kai and Gojcic, Zan and Gilitschenski, Igor and Fidler, Sanja and Vijaykumar, Nandita and Wang, Zian},
  journal={arXiv preprint arXiv:2509.03680},
  year={2025}
}