InstantHDR: Single-forward Gaussian Splatting Initialization for HDR 3D Reconstruction

Dingqiang Ye, Jiacong Xu, Jianglu Ping, Yuxiang Guo, Chao Fan, and Vishal M. Patel

ECCV 2026 · Official Implementation

Project Page | Paper | Checkpoint | Dataset

InstantHDR reconstructs an HDR Gaussian scene from multi-exposure LDR images in a single forward pass, enabling novel-view rendering at different exposures.

Requires Linux, Python 3.10, CUDA 12.1, and an NVIDIA Ampere or newer GPU. Run all commands from the repository root.

conda create -n instanthdr python=3.10 -y
conda activate instanthdr
pip install torch==2.2.0 torchvision==0.17.0 --index-url https://download.pytorch.org/whl/cu121
pip install -r requirements.txt
pip install xformers==0.0.24
pip install torch-scatter==2.1.2+pt22cu121 -f https://data.pyg.org/whl/torch-2.2.0+cu121.html
pip install https://github.com/nerfstudio-project/gsplat/releases/download/v1.4.0/gsplat-1.4.0%2Bpt22cu121-cp310-cp310-linux_x86_64.whl

The pretrained InstantHDR checkpoint is hosted in the Hugging Face dataset repository Bugjudger/HDR_Pretrain, at checkpoints/InstantHDR.ckpt. Download it to checkpoints/InstantHDR.ckpt in this project:

mkdir -p checkpoints
curl -L --fail https://huggingface.co/datasets/Bugjudger/HDR_Pretrain/resolve/main/checkpoints/InstantHDR.ckpt \
  -o checkpoints/InstantHDR.ckpt
export INSTANTHDR_CHECKPOINT="$PWD/checkpoints/InstantHDR.ckpt"

python demo.py

Open http://127.0.0.1:7860. Try Bear, Chair, or Dog with prefilled exposure values, or upload your own overlapping images and enter their log2 exposure times. Bear is loaded by default.

1. Reconstruct — Use at least two input views and click 1. Reconstruct to create the HDR scene.
2. Render — Adjust the output exposure and click 2. Render at selected exposure to explore the scene at different brightness levels.

Run inference.py on the four included Bear views. After completing Installation and downloading the checkpoint, run the following from the repository root:

conda activate instanthdr
python inference.py \
  --checkpoint checkpoints/InstantHDR.ckpt \
  --images examples/bear/images \
  --exposures examples/bear/exposure.json \
  --output outputs/bear \
  --render-ev 0

The checkpoint is passed explicitly; INSTANTHDR_CHECKPOINT is not needed for this command. The exposure JSON maps filenames to positive linear exposure times. Bear uses 2, 0.125, 32, and 0.125 for its four input views. --render-ev 0 selects an output log2 exposure of 0 (linear exposure 1).

Outputs are saved to:

outputs/bear/
├── rgb.mp4       # Novel-view RGB video
├── depth.mp4     # Depth visualization video
├── scene.pt      # HDR Gaussian tensors and cameras
└── preview.ply   # Gaussian PLY preview

To try another included scene, replace both examples/bear paths with examples/chair or examples/dog, and choose a matching output directory.

Download HDR-Pretrain and follow the data preparation instructions. Run from the repository root, in the environment prepared during Installation.

Set DATASET_ROOT to the directory containing scenes/, train_index.json, and test_index.json, not to scenes/ itself. The training index includes all 168 scenes; the validation index selects five of them and uses their test views.

DATASET_ROOT=/path/to/HDR-Pretrain
export TMPDIR=/tmp  # Keep data-loader temporary files on local storage.
CUDA_VISIBLE_DEVICES=0 python -m src.main +experiment=hdr \
  "dataset.hdr.roots=[$DATASET_ROOT]" \
  trainer.max_steps=10000 \
  model.encoder.pretrained_weights=checkpoints/anysplat.safetensors  # Download: https://huggingface.co/lhjiang/anysplat/resolve/main/model.safetensors

Choose an available GPU with CUDA_VISIBLE_DEVICES. Setting TMPDIR=/tmp avoids worker cleanup errors when the default temporary directory is on NFS.

Training uses Hydra to combine the following files. Command-line overrides take precedence over the YAML settings.

Edit config/experiment/hdr.yaml to adjust training settings. Results are saved to output/exp_hdr/. To resume, add checkpointing.load=/path/to/training.ckpt with trainer.max_steps greater than the saved step.

Refine an InstantHDR scene using multi-exposure LDR images. The pipeline first refines camera poses for 100 iterations, then optimizes the Gaussians, tone-mapping parameters, and training-camera poses.

CUDA_VISIBLE_DEVICES=0 python -m src.post_opt.simple_trainer_hdr default \
  --checkpoint checkpoints/InstantHDR.ckpt \
  --data-dir /path/to/scene/images \
  --result-dir outputs/post_opt/example \
  --eval-mode ldr \
  --max-steps 1000

Provide train*/test* images and exposure.json; see data format. Post optimization requires at least two training views and one test view; the demo presets contain only training views. Use one GPU per scene. Results include checkpoints, renders, and metrics.

Both training and test images participate in camera estimation and initial pose refinement; only training images supervise Gaussian optimization. hdr in the script name refers to the scene representation. To evaluate HDR output, use --eval-mode hdr with matching EXR references; optimization still uses LDR supervision.

HDR-Pretrain is a synthetic pretraining dataset for 3D HDR reconstruction, designed to train feed-forward models that reconstruct HDR 3D scenes from multi-view, multi-exposure LDR images. It contains 168 indoor scenes, built from HSSD assets and rendered with Blender Cycles. Each scene provides:

• 35 viewpoints sampled on a 5 × 7 grid, rendered at 448 × 448 resolution.
• Five exposure levels per view, paired with 32-bit HDR ground truth.
• Depth and normal maps, with one of AgX, Filmic, or Standard tone-mapping operators selected per scene.

Download the dataset from Hugging Face. See the data preparation instructions for the directory layout, exposure metadata, and training/validation index format, then follow Training to use it.

The Blender scene files and dataset generation code are also available on Hugging Face: rendering script (instanthdr_render.py) and batch launcher (instanthdr.sh).

The script loads each .blend scene, samples a 5 × 7 grid of camera views, and uses Blender Cycles to render multi-exposure LDR images, HDR references, depth, and normals. It also exports camera parameters in COLMAP format and writes exposure.json. To generate data, download the source scenes into blender_scenes/ and keep the scripts at the dataset root. Install Blender 4.5.6, then run BLENDER=/path/to/blender CUDA_VISIBLE_DEVICES=0 bash instanthdr.sh; outputs go to scenes/<scene_id>/. Exposure levels, resolution, and tone-mapping choices can be adjusted in the rendering script's CONFIG dictionary.

If you find our work useful, please consider citing:

@inproceedings{ye2026instanthdr,
  title={InstantHDR: Single-forward Gaussian Splatting Initialization for HDR 3D Reconstruction},
  author={Ye, Dingqiang and Xu, Jiacong and Ping, Jianglu and Guo, Yuxiang and Fan, Chao and Patel, Vishal M.},
  booktitle={European Conference on Computer Vision (ECCV)},
  year={2026}
}

Built on AnySplat, VGGT, DUSt3R, CroCo, and gsplat. See third-party licenses.