ECCV 2026

DualDiff3D: Dual Structure-Appearance Diffusion Priors for Reliability-Enhanced 3D Gaussian Splatting

Decoupling structure and appearance for sharper, more reliable sparse-view 3D reconstruction.

Qian Wang^* Yu Wang^* Weiqi Li Xinhua Cheng Xiandong Meng Ronggang Wang Jian Zhang^†

^* Equal contribution ^† Corresponding author

Peking University Pengcheng Laboratory Guangdong Provincial Key Laboratory of UHD Immersive Media Technology

Code arXiv View videos

Video comparison

Explore the reconstruction results.

Compare synchronized novel-view renderings from 3DGS, DIFIX3D, and DualDiff3D.

Baseline3DGS

Single priorDIFIX3D

Ours

Dual prior + reliable RRODualDiff3D

All three videos are synchronized. Click any video or use the control to pause and inspect details.

Overview

One prior for structure. One prior for appearance.

Sparse input views leave 3D Gaussian Splatting with incomplete geometry and conspicuous artifacts. Existing diffusion-based refiners mix rendered and reference views in one network, despite their fundamentally different roles: a rendered view defines structure, while reference views provide appearance.

We introduce DualDiff, a two-branch refinement pipeline connected by Structure-Appearance Attention. A dedicated structure branch preserves the viewpoint, layout, and geometric correspondence of low-quality novel views, while an appearance branch supplies textures, colors, and details from reference images. Structural queries attend to both structural and appearance features, introducing reference guidance without sacrificing viewpoint consistency.

Building on this refinement model, DualDiff3D uses a reliability-enhanced Render-Refine-Optimize loop to turn refined novel views into stronger 3D supervision. Progressive sampling and filtering selects useful nearby viewpoints, confidence-driven weighting evaluates reliability at the pixel level, and validation with rollback prevents unstable updates from degrading the reconstruction.

Method

Separate first. Fuse with purpose.

DualDiff resolves the conflict between viewpoint-dependent geometry and view-consistent appearance.

Motivation comparing single-branch DIFIX with the dual-branch DualDiff design — **Motivation.** A single branch blurs as the reference viewpoint moves farther away. Decoupling structure and appearance avoids this conflict.

DualDiff pipeline with structure and appearance diffusion branches connected by Structure-Appearance Attention — **DualDiff.** The structure branch takes the rendered novel view to preserve viewpoint and geometry, while the appearance branch extracts texture and detail from reference views. Structure-Appearance Attention uses structural hidden states as queries and combines structural and appearance features as keys and values.

DualDiff3D reliability-enhanced Render-Refine-Optimize loop — **DualDiff3D.** The Render-Refine-Optimize loop progressively augments 3DGS with refined novel views. Progressive Sampling and Filtering selects reliable viewpoints, Confidence-Driven Weighting evaluates pixel-level reliability, and validation with rollback prevents unstable updates.

Results

Quantitative and qualitative comparisons.

Visual comparison of novel-view refinement by DIFIX and DualDiff across multiple scenes — DualDiff refinement DualDiff preserves the geometry of rendered novel views while transferring fine appearance details from reference views.

DL3DV · 3 views+1.20 dB

PSNR over DIFIX3D

LLFF · 3 views+1.13 dB

PSNR over DIFIX3D

Refinement speed600 ms

per image on RTX 4090

Across all settings21 / 21

best DL3DV & LLFF core metrics

Qualitative reconstruction comparison on DL3DV and LLFF datasets — DualDiff3D reconstruction DualDiff3D improves photorealism and geometric coherence across both in-domain DL3DV and out-of-domain LLFF scenes.

Citation

Build on DualDiff3D.

If this work is useful for your research, please cite our ECCV 2026 paper.

@inproceedings{wang2026dualdiff3d,
  title     = {DualDiff3D: Dual Structure-Appearance Diffusion
               Priors for Reliability-Enhanced 3D Gaussian Splatting},
  author    = {Qian Wang and Yu Wang and Weiqi Li and Xinhua Cheng
               and Xiandong Meng and Ronggang Wang and Jian Zhang},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2026},
  month     = {jun}
}