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

ECCV 2026

Dingqiang Ye1,* Jiacong Xu1,* Jianglu Ping1 Yuxiang Guo1 Chao Fan2 Vishal M. Patel1

1 Johns Hopkins University2 Shenzhen University* Equal contribution

Paper Code Dataset Poster
Paper Figure 1: four-view HDR reconstruction and rendering comparison between GaussHDR, AnySplat, and InstantHDR.
InstantHDR reconstructs HDR scenes from four multi-exposure views in a single forward pass.

Abstract

High dynamic range (HDR) novel view synthesis (NVS) aims to reconstruct HDR scenes from multi-exposure low dynamic range (LDR) images. Existing HDR pipelines heavily rely on known camera poses, well-initialized dense point clouds, and time-consuming per-scene optimization. Current feed-forward alternatives overlook the HDR problem by assuming exposure-invariant appearance. To bridge this gap, we propose InstantHDR, a feed-forward network that initializes 3D HDR scenes from uncalibrated multi-exposure LDR collections in a fast single forward pass. Specifically, we design a geometry-guided appearance modeling for multi-exposure fusion, and a meta-network for generalizable scene-specific tone mapping. Due to the lack of HDR scene data, we build a pre-training dataset, called HDR-Pretrain, for generalizable feed-forward HDR models, featuring 168 Blender-rendered scenes, diverse lighting types, and multiple camera response functions. Comprehensive experiments show that our InstantHDR delivers a single-forward HDR initialization at ∼700× the speed of SoTA optimization-based methods, and reaches comparable quality in real settings after lightweight post-optimization while remaining ∼20× faster. All code, models, and datasets: https://github.com/Bugjudger/InstantHDR.

Method Overview

Original InstantHDR method diagram: geometry-guided appearance modeling and scene-specific tone mapping.

Geometry-guided appearance fusion. We reuse attention from the frozen geometry branch to find cross-view correspondences and fuse complementary exposure information. Difference of Gaussians restores fine pixel-level details lost during patch-level fusion.

Scene-specific tone mapping. A meta-network predicts a lightweight tone mapper for each scene, adapting to camera response differences and enabling exposure-controlled rendering.

HDR pretraining. HDR-Pretrain provides 168 synthetic scenes with varied lighting and camera responses, supporting generalizable feed-forward HDR reconstruction.

Demo Videos

Novel-view rendering on Bear, Box, and Chair, comparing GaussHDR, AnySplat, and InstantHDR.

Bear

Input: Four multi-view LDR images captured at different exposures, without camera poses or an initial point cloud.

Bear multi-exposure input 1Bear multi-exposure input 2Bear multi-exposure input 3Bear multi-exposure input 4
GaussHDR
AnySplat
InstantHDR (Ours)

Box

Input: Four multi-view LDR images captured at different exposures, without camera poses or an initial point cloud.

Box multi-exposure input 1Box multi-exposure input 2Box multi-exposure input 3Box multi-exposure input 4
GaussHDR
AnySplat
InstantHDR (Ours)

Chair

Input: Multi-view LDR images captured at different exposures, without camera poses or an initial point cloud.

Chair multi-exposure input 1Chair multi-exposure input 2Chair multi-exposure input 3Chair multi-exposure input 4Chair multi-exposure input 5Chair multi-exposure input 6Chair multi-exposure input 7Chair multi-exposure input 8
GaussHDR
AnySplat
InstantHDR (Ours)

Ablations & Analysis

Why each component matters

Ground truth, full InstantHDR, and variants without cross-view attention, exposure normalization, or upsampling.
Removing cross-view attention introduces ghosting; removing exposure normalization shifts brightness; removing upsampling blurs fine details.

Correspondence across views and exposures

Geometry attention maps matching leaf, doorframe, cup, and mirror patches across viewpoints and exposures.
The frozen geometry encoder matches corresponding patches across large viewpoint and exposure changes, guiding appearance fusion.

Adapting tone mapping to each scene

Learned monotonic camera-response curves for the Box, Bear, and Dog scenes.
MetaNet learns distinct, monotonic response curves for different scenes; Box saturates earlier than the other examples.

HDR-Pretrain Dataset

168 photorealistic indoor scenes rendered in Blender for generalizable HDR reconstruction.

HDR-Pretrain examples: multi-exposure LDR images, HDR ground truth, normal and depth maps, with Standard, AgX, and Filmic tone mapping.
Each viewpoint includes five exposure-bracketed LDR images, 32-bit HDR ground truth, and depth and normal maps; Standard, AgX, and Filmic tone mapping provide diverse camera responses.
Download on Hugging Face

BibTeX

@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 = {Computer Vision -- ECCV 2026},
  year      = {2026},
  publisher = {Springer}
}