Geo-ID: Test-Time Geometric
Consensus for Cross-View Consistent Intrinsics

ECCV 2026

A training-free, inference-time framework that turns off-the-shelf single-view intrinsic predictors into cross-view consistent ones — coupling independent per-view predictions through sparse geometric correspondences.

Alara Dirik¹ · Stefanos Zafeiriou¹

¹Imperial College London

Paper Code soon

Geo-ID teaser: inconsistent vs consistent intrinsics across views — Existing single-view predictors estimate PBR parameters independently per view, producing **cross-view inconsistencies**. Geo-ID couples per-view predictions through geometric correspondences at test time, yielding consistent decompositions that enable coherent material editing and relighting in downstream neural scene representations.

Contributions

— i

Training-free consistency

An inference-time framework that generates cross-view consistent decompositions from sparse, unordered image collections by coupling frozen single-view diffusion predictors.

— ii

Geometry-driven consensus

A consensus formulation that aggregates per-view intrinsics across sparse 3D correspondences using robust, confidence-weighted voxel statistics.

— iii

Consensus-guided diffusion

A sampling strategy that injects sparse cross-view constraints into the denoising process, preserving the generative prior of the base model while enforcing agreement.

Abstract

Intrinsic image decomposition aims to estimate physically based rendering (PBR) parameters such as albedo, roughness, and metallicity from images. While recent methods achieve strong single-view predictions, applying them independently to multiple views of the same scene often yields inconsistent estimates, limiting their use in downstream applications such as editable neural scenes and 3D reconstruction. Video-based models can improve cross-frame consistency but require dense, ordered sequences and substantial compute, limiting their applicability to sparse, unordered image collections. We propose Geo-ID, a novel test-time framework that repurposes pretrained single-view intrinsic predictors to produce cross-view consistent decompositions by coupling independent per-view predictions through sparse geometric correspondences that form uncertainty-aware consensus targets. Geo-ID is model-agnostic, requires no retraining or inverse rendering, and applies directly to off-the-shelf intrinsic predictors. Experiments on synthetic benchmarks and real-world scenes demonstrate substantial improvements in cross-view intrinsic consistency as the number of views increases, while maintaining comparable single-view decomposition performance. We further show that the resulting consistent intrinsics enable coherent appearance editing and relighting in downstream neural scene representations.

Method

Geo-ID pipeline overview — **Overview of Geo-ID.** A pretrained geometry transformer (VGGT) predicts cameras and confidence-weighted 3D point maps. A frozen single-view diffusion model produces independent per-view intrinsics; high-confidence 3D points are voxelised and aggregated into a robust voxel-level consensus that is reprojected into each view. A second, consensus-guided diffusion pass injects these targets as sparse guidance at selected denoising steps to generate cross-view consistent predictions.

Phase 1

Geometry-guided correspondences

VGGT predicts camera parameters and dense world-frame point clouds with per-pixel confidence from an unordered image set. Only high-confidence points are retained.

Phase 2

Voxel-based consensus

Points are pooled and partitioned into voxels. Per-voxel intrinsics are fused with a confidence-weighted median; a robust dispersion estimate rejects outliers and weights guidance.

Phase 3

Consensus-guided sampling

Reprojected targets steer a second per-view diffusion pass via a Huber consensus loss on the latent, applied to the last 80% of steps. Weights stay frozen; sampling stays per-view.

Projected consensus targets visualization — **What the guidance signal looks like.** For representative scenes we show the input view, the projected albedo consensus, and the inverse-dispersion confidence (brighter = more confident). Indoor scenes yield denser, higher-confidence targets than outdoor scenes — yet even sparsely covered regions improve, since the consensus gradient propagates spatially through the U-Net beyond the exact target locations.

Quantitative Results

**Cross-view intrinsic consistency** — mean per-correspondence MAD (lower is better). Geo-ID reduces cross-view disagreement across every modality and dataset, with the largest gains on roughness and metallicity. Base models and Geo-ID compared at 32 views.
Method	Views	MipNeRF-360 Indoor			MipNeRF-360 Outdoor			Tanks & Temples
Method	Views	Alb.	Rou.	Met.	Alb.	Rou.	Met.	Alb.	Rou.	Met.
IDArb	16	0.258	0.316	0.289	0.224	0.310	0.280	0.256	0.309	0.299
Colorful Diffuse	32	0.101	–	–	0.070	–	–	0.103	–	–
Diffusion Renderer (ordered)	32	0.068	0.038	0.043	0.043	0.041	0.019	0.060	0.040	0.028
Diffusion Renderer (unordered)	32	0.104	0.075	0.087	0.089	0.100	0.102	0.110	0.100	0.129
RGB↔X	32	0.114	0.100	0.206	0.070	0.144	0.192	0.109	0.105	0.219
RGB↔X + Geo-ID	32	0.098	0.065	0.114	0.054	0.079	0.085	0.097	0.076	0.166
Marigold Appr.	32	0.091	0.096	0.111	0.073	0.080	0.070	0.098	0.100	0.118
Marigold Appr. + Geo-ID	32	0.076	0.082	0.100	0.061	0.062	0.044	0.082	0.080	0.095

Consistency improves monotonically with the number of input views (V ∈ {4, 8, 16, 32}); gains are confirmed on InteriorVerse using ground-truth correspondences, ruling out alignment with VGGT's geometric biases.

**Per-view decomposition quality** on InteriorVerse — enforcing cross-view consistency does not degrade single-view accuracy, and in places slightly improves it.
Method	Albedo PSNR↑	SSIM↑	LPIPS↓	Metal. RMSE↓	Rough. RMSE↓
Diffusion Renderer (video)	21.9	0.87	0.17	0.28	0.35
RGB↔X	16.4	0.78	0.19	0.44	0.38
RGB↔X + Geo-ID	16.4	0.78	0.19	0.42	0.35
Marigold Appr.	19.5	0.85	0.19	0.20	0.25
Marigold Appr. + Geo-ID	19.5	0.86	0.19	0.19	0.23

Cross-View Consistency

Each row below shows a single modality across all input views of a scene at once. The left state is the base predictor run independently per view; drag right to reveal Geo-ID. Watch the whole strip settle: per-view color drift and material flicker on shared surfaces disappear, while fine detail is preserved.

◄► Drag any strip — Base to Geo-ID across every view

Indoor — MipNeRF-360 RGB↔X → + Geo-ID

Input views 4 views of the same scene

Input

Albedo Base ► Geo-ID

Base Geo-ID

Roughness Base ► Geo-ID

Base Geo-ID

Metallicity Base ► Geo-ID

Base Geo-ID

Outdoor — Tanks & Temples RGB↔X → + Geo-ID

Input views 4 views of the same scene

Input

Albedo Base ► Geo-ID

Base Geo-ID

Roughness Base ► Geo-ID

Base Geo-ID

Metallicity Base ► Geo-ID

Base Geo-ID

Comparison with single-view, multi-view and video baselines — **Against single-view, multi-view & video baselines** on MipNeRF-360 Garden. Single-view methods (RGB↔X, IDArb) are detailed but cross-view inconsistent; video methods (Diffusion Renderer) are consistent but bake in lighting. Geo-ID keeps single-view detail while achieving cross-view consistency.

Downstream Applications

Relighting and material editing from Geo-ID intrinsics — **Relighting & material editing.** We train a MeshSplatting neural scene representation on Geo-ID albedo, then render the reconstructed meshes under novel HDR environment maps (top) and edit segmented material regions (bottom). Because Geo-ID makes albedo view-independent, edits and relighting propagate coherently across views — without the seams and color discontinuities that inconsistent per-view estimates would bake in.

BibTeX

@inproceedings{dirik2026geoid,
  title     = {Geo-ID: Test-Time Geometric Consensus for Cross-View Consistent Intrinsics},
  author    = {Dirik, Alara and Zafeiriou, Stefanos},
  booktitle = {Proceedings of the European Conference on Computer Vision (ECCV)},
  year      = {2026}
}