⚡ FlashRelight

Portrait Video Relighting with Dynamic Lighting
NeurIPS 2026

Jiahui Sheng^1,*, Le Li^2,†, Limin Lin², Jiahao Li², Wansen Feng², Hong Gu², Shuhan Chen¹, Xiaorun Li^1,‡, Rui Hu^2,‡
1 Zhejiang University, China · ² vivo BlueImage Lab, vivo Mobile Communication Co., Ltd., China
_{* Work done during an internship at vivo.   † Project leader.   ‡ Corresponding authors.}

FlashRelight relights portrait videos with an HDR environment map or transfers lighting from a reference video. Its first stage, FlashDelight, can also run on its own to suppress the original lighting.

🧩 Method

FlashDelight estimates a neutrally lit portrait video. FlashRelight uses that estimate and a frame from the input video to preserve the subject’s appearance while applying the target lighting.

🗂️ Training data

We render 24,000 video pairs from static one-light-at-a-time (OLAT) captures using known lighting sequences. Over 10,000 generated video pairs show subjects moving under changing light.

How the training pairs differ between the two stages

For FlashDelight, we add synthetic lighting changes to generated, neutrally lit clips. For FlashRelight, we generate clips under changing light and use FlashDelight to produce their neutral counterparts.

🎬 Examples

Delighting

Each pair shows the original video on the left and the FlashDelight output on the right.

Relighting

Each comparison below shows the input portrait, lighting control, delighted portrait, and relit result, from left to right.

Lighting control	Watch the comparison
HDR sequence	Light moving around the portrait
Single HDR panorama (fixed or rotating)	Meadow 2 rotating at 60°/s
Reference video	Studio light sweeping across the face
Text prompt	Dappled light moving across the face

Compare reference lighting across portraits

The top row shows the reference lights. Each row below shows an input portrait on the left and its relit results under each reference.

🚀 Quick start

1. Install

Use Linux, Conda, and an NVIDIA GPU. The default environment uses Python 3.10 and PyTorch 2.1.0 with CUDA 12.1, as tested on the H20.

git clone https://github.com/vivoCameraResearch/FlashRelight.git
cd FlashRelight
bash install.sh
conda activate flashrelight

For the RTX 5090 D examples, we used Python 3.11 and PyTorch 2.14.0+cu132. If you already have a PyTorch environment for your GPU, install the remaining packages with pip install -r requirements.txt.

2. Weights

Download the FlashDelight and FlashRelight checkpoints together with the shared Wan2.1 files. Hugging Face weights

python scripts/download_weights.py --dest weights --skip_matting_model

The bundled portraits already have black backgrounds. For footage that needs background removal, omit --skip_matting_model during setup and use --matting birefnet at inference. Offline setup and weight layout

3. Relight

python scripts/infer.py --source assets/sources/generated_gesture.mp4 --matting none \
  --mode env --env_map assets/envmaps/04_random_great_circle

Open outputs/generated_gesture_env_relit.mp4 for the relit portrait (49 frames, 480 × 832 pixels, 16 fps). The same folder contains the delighted portrait (*_delit.mp4) and the comparison (*_compare.mp4).

For your own input, pass a portrait video or still image to --source. All six bundled environment maps retain their original HDR values. FlashRelight applies the 0.1 scale automatically during inference.

Run all four examples with python scripts/infer.py --samples assets/samples.json. Sample configurations

Use a single panorama, a reference video, or a text prompt

Single HDR panorama. A panorama stays fixed for all 49 frames by default. Set --env_rotation_speed in degrees per second to rotate it. Positive values move the light to the right in the panorama preview.

python scripts/infer.py --source assets/sources/generated_portrait.mp4 --matting none \
  --mode env --env_map assets/envmaps/meadow_2_4k.exr --env_rotation_speed 60

The --env_map flag accepts a single EXR/HDR panorama, a directory of ordered EXR/HDR frames, or a NumPy array such as [T,3,H,W]. The model resizes each map to 128 × 64 pixels.

Reference video. Transfer the lighting from the studio clip to a different portrait.

python scripts/infer.py --source assets/sources/generated_gesture.mp4 --matting none \
  --mode light_ref --light_ref assets/light_refs/studio_sweep.mp4

Text prompt. Text control works best with lighting descriptions close to the training captions. Example prompts

python scripts/infer.py --source assets/sources/generated_portrait.mp4 --matting none \
  --mode text --prompt "Across the face, irregular dappled light patches wander with an organic motion."

🌓 Standalone FlashDelight

To run delighting alone, use the FlashDelight checkpoint and the shared Wan2.1 files. If you already prepared both stages, go straight to the inference command.

python scripts/download_weights.py --dest weights --delight_only --skip_matting_model

python scripts/infer.py --delight_only --source assets/sources/generated_portrait.mp4 --matting none

This writes outputs/generated_portrait_delit.mp4. For the portrait with glasses, use --source assets/sources/generated_gesture.mp4. See both delighting results

Use FlashDelight from Python

from flashrelight import FlashRelightPipeline
from flashrelight.utils import video_io

pipe = FlashRelightPipeline("weights", delight_only=True)
frames = pipe.preprocess("assets/sources/generated_portrait.mp4", matting="none")
video_io.save_mp4(pipe.delight_frames(frames), "delighted.mp4", fps=16)

🎛️ Gradio demo

Gradio lets you run FlashRelight from a browser. After preparing the weights, start the web interface on your GPU machine.

python app.py --matting none

Open the local URL printed in the terminal. Choose a bundled portrait or upload your own, select a lighting control, and click Relight. For a delighting-only interface, run python app.py --delight_only --matting none.

🙏 Acknowledgements

FlashRelight uses the DiT, 3D VAE, and umT5 text encoder from Wan2.1. We use BiRefNet for optional portrait matting. The examples include HDR maps from Poly Haven (CC0).

Example asset credits · Third-party software

📖 Citation

@misc{sheng2026flashrelightportraitvideorelighting,
      title={FlashRelight: Portrait Video Relighting with Dynamic Lighting},
      author={Jiahui Sheng and Le Li and Limin Lin and Jiahao Li and Wansen Feng and Hong Gu and Shuhan Chen and Xiaorun Li and Rui Hu},
      year={2026},
      eprint={2611.06355},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2611.06355},
}