An arXiv preprint describes a method for turning multi-view images into editable 3D assets that can be rendered under changed lighting. It asks whether a staged framework can reconstruct relightable UV-PBR mesh assets directly from those images. Here, a UV-PBR asset means a mesh whose surface properties are stored in maps tied to its UV coordinates, so those properties can be edited and the object relit.
ExMesh++ divides the process into two stages. Stage I jointly optimizes mesh geometry, topology and an RGB texture, updating the UV layout as vertices are split or merged. Stage II freezes the reconstructed mesh, topology and UV coordinates, then optimizes maps for base color, roughness, normals and optional metallic values alongside environment lighting. The same shared UV-space maps are queried at both primary surface points and secondary hit points in the indirect-light calculation.
The tests covered three object sets
The evaluation used 15 real-world objects from DTU, four CAD objects from Synthetic4Relight, and 14 real objects from Stanford-ORB under seven real environments. NeRF-Synthetic was used only for Stage-I ablations. The reported measures covered Chamfer Distance for geometry, PSNR, SSIM and LPIPS for rendered images, mean squared error for roughness, plus runtime, vertex count, depth error and normal error.
The geometry result was a tradeoff
On DTU, ExMesh++ reported an average Chamfer Distance of 0.58 and 13 minutes of training. PGSR and GeoSVR had lower reported Chamfer Distance, but they took longer to train and used larger geometric representations. The result is therefore a compact and faster reported reconstruction on that benchmark, even though it did not lead on the geometry-distance measure.
Synthetic4Relight produced the paper's highest listed relighting PSNR among the compared methods. Its novel-view scores were PSNR 35.76, SSIM 0.971 and LPIPS 0.053. Relighting scores were 34.19, 0.962 and 0.061, while albedo scores were 30.16, 0.953 and 0.068. Roughness mean squared error was 0.005, and the reported runtime was 30 minutes.
Stanford-ORB tested novel views and novel-scene relighting on real objects. For novel views, ExMesh++ reported PSNR-H 31.27, PSNR-L 39.52, SSIM 0.990 and LPIPS 0.014. For novel-scene relighting, it reported PSNR-H 26.60, PSNR-L 34.03, SSIM 0.980 and LPIPS 0.022. Depth error was 0.30, normal error 0.02 and Chamfer Distance 0.30, with a 30-minute runtime. The paper presents these as the strongest listed novel-scene relighting metrics and lowest listed geometry errors in that comparison.
Topology and indirect light mattered
The topology ablation compared different ways of changing the mesh. In the NeRF-Synthetic Stage-I test, the full strategy reported Chamfer Distance 0.64, PSNR 29.32 and 100,000 vertices. A version using only merging reported 1.81, 23.51 and 6,000 vertices, while the random variants underperformed the full strategy. Across those variants, the full strategy had the strongest reported combined geometry and rendering result, at the cost of a larger mesh than the merge-only version.
Material and lighting ablations also favored the full setting. It reported relighting PSNR 34.19 and albedo PSNR 30.16, compared with 33.72 and 29.96 for Direct Only. The paper reports the full shared-material, indirect-lighting configuration as best overall after querying shared UV-PBR materials at secondary hit points. In a qualitative DTU comparison, ExMesh++ showed more continuous, environment-dependent indirect-light variation around a can lid and in occluded regions than the compared methods.
Richer lighting came at a cost
That extra light transport came with a measurable cost in the reported implementation. The full pipeline took 30 minutes per object, including about 3 minutes for initialization, 10 minutes for Stage I and 17 minutes for Stage II. With indirect illumination disabled, the total fell to 23 minutes. Rendering was 15.09 frames per second for the full model and 40.47 FPS in direct-only mode. Timing and frame rate depend on implementation and hardware.
The model has clear boundaries
The appearance model is an opaque, isotropic metallic-roughness BRDF, meaning it represents a particular class of surface reflection. It does not cover anisotropic reflection, transmission or subsurface scattering. Metallic optimization was disabled in the benchmark experiments, and metallic-map recovery was not quantitatively evaluated.
The lighting model is similarly limited: it includes one-bounce diffuse interreflection but excludes multi-bounce transport and indirect specular reflection. Periodic UV regeneration uses CPU-based xatlas and may bottleneck high-resolution meshes. The paper leaves larger scenes and memory management unresolved, so the tests do not establish scalability to large multi-object scenes or complex production assets.
An editable asset, within limits
The paper demonstrates the output in Blender through relighting, composition with artist-created assets, and direct texture-space edits to surface paint, base color, roughness and metallic channels. These examples show the assets working in the demonstrated digital-content-creation workflow, but no independent user study or production-scale evaluation was reported.
A benchmark result, not a universal verdict
The work remains a preprint, and its evidence comes from benchmark comparisons. The document identifies itself as arXiv:2608.24109v2, dated 29 August 2026. The comparisons come without confidence intervals or significance tests, so the reported advantages describe the tested objects and environments rather than a guaranteed result elsewhere.
On those tests, ExMesh++ offers a route to editable, relightable UV-PBR assets from multi-view images. The results do not establish universal superiority, quantitative metallic recovery, complete light transport or scalability beyond the tested object-level settings.
Paper data and sources
Original title: ExMesh++: From Multi-View Images to Relightable UV-PBR Mesh Assets via Topology-Adaptive Reconstruction and Decomposition
Authors: Chuanjin Fan, Lifan Wu, Wenjie Chang et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: Not available
Original paper · Full text