This is a PyTorch implementation of the following paper:

WildCap: Facial Albedo Capture in the Wild via Hybrid Inverse Rendering, CVPR 2026.

Yuxuan Han, Xin Ming, Tianxiao Li, Zhuofan Shen, Qixuan Zhang, Lan Xu, and Feng Xu

Project Page | Video | Paper

Unlike my previous projects (i.e., CoRA and DoRA), achieving 4K-level high-quality facial appearance capture no longer requires data capture in a darkroom. WildCap can reconstruct even higher-quality facial reflectance maps than DoRA from a smartphone video of human face captured in-the-wild!

Technically, the biggest advantage of WildCap over previous works (e.g., FLARE, NextFace, and Xu et al.) lies in its robustness to hard shadows. Even when there are large hard shadows on the face — which are common in in-the-wild scenarios, as both indoor lights and the sun can cast such shadows — WildCap still produces clean reflectance maps, a capability that previous methods cannot achieve.

Step 1: Set up the code environment according to ENV.md.

Step 2: Prepare the data. You can either run our preprocessed dataset (download and put it into data/wildcap_dataset), or process videos captured by yourself into our data format following PROCESS.md.

Step 3: Running the appearance capture code:

# this script will run all the subjects in our preprocessed dataset
sh run.sh  

# run an specific subject (all the hyper-parameters are shared across subjects)
python run_wildcap.py \
    --log_dir data/ljf/results/debug_env \
    --pho_scale 1.0 \
    --grad_scale 1.0 \
    --dps_steps 1000 \
    --uv_ds 4 \
    --vis_freq 100 \
    --env_light_path xxx \
    --eta 1.0 \
    --baseline 0 \
    --reg_scale 1.0 \
    --tv_scale 0.1 \
    --n_dpps 20 \
    --sr_scale 1 \
    --shadow_mask_path data/ljf/shadow_mask.png \
    --shade_init 1 \
    --sche_light 1 \
    --main_light_type splitsum \
    --opt_light_iter 300 \
    --light_res 96 \
    --init 1 \
    --init_add_noise_step 600 \
    --data_root data/ljf \
    --texture_path data/ljf/uv_switchlight.png \
    --init_map_root data/ljf/initmap \
    --model_dir models

Step 4: SR the reflectance maps to 4K resolution (TODO)

Step 5: Full-head texture and geometry completion (for visualization usage) (TODO)

Compared with the original paper, this open-source implementation includes the following key improvements:

1. Better posterior sampling strategy: We adopt DPPS in this code release, which is an enhanced version of the DPS technique mentioned in the paper.

2. Better base lighting model: We use Split-Sum to model the base lighting, and we find it performs better than SH lighting mentioned in the paper.

3. Open-source compatible reflectance initialization strategy: The reflectance initialization strategy proposed in the paper relies on the closed-source Light Stage scan dataset. To maintain the open-source nature of this project, we apply a new strategy: everyone uses the same initial specular and normal maps, while the initial diffuse map is a pure-color image. Compared with the paper version, this strategy avoids copyright issues and achieves almost the same performance as the paper’s method. See create_init_map.py and PROCESS.md for more details.

If you have any questions or are interested in collaboration, please contact Yuxuan Han (hanyx22@mails.tsinghua.edu.cn).

Please include the following citations if it helps your research:

@inproceedings{han2026wildcap,
    author = {Han, Yuxuan and Ming, Xin and Li, Tianxiao and Shen, Zhuofan and Zhang, Qixuan and Xu, Lan and Xu, Feng},
    title = {WildCap: Facial Albedo Capture in the Wild via Hybrid Inverse Rendering},
    journal={CVPR},
    year={2026}
}