Recent generation methods produce 3D assets with detailed geometry and rich textures. However, these assets are non-emissive: they reflect light but emit none. We study a new task, emission generation: predicting spatially varying RGB emission over a 3D asset's surface from its geometry, non-emissive materials, and a rendered view. To support this task, we introduce LightGenBench. Our benchmark makes three contributions. First, we curate a dataset of 37k artist-authored emissive assets in UV, voxel, and multi-view form. Second, we establish baselines across four method families: UV-domain, sparse-voxel, multi-view, and segmentation-based generation. Third, we design an evaluation protocol that measures where an asset emits and how, complemented by user-study comparisons on renderings with varying emission strength. Across the four families, the sparse-voxel baseline TRELLIS.2-Emission performs best because generating emission natively in 3D avoids the failure modes of the other representations, such as the fracturing artifacts of UV space and the occluded emitters that multi-view generation cannot see. However, TRELLIS.2-Emission only reaches an IoU of 0.271, leaving the task far from solved. We hope LightGenBench establishes emission generation as a benchmarkable problem for 3D generation.
Figures & tables
Figure 2: Samples from our curated dataset. Each pair shows an object under standard lighting (top) and the same object rendered in a dark room, lit only by its own emission (bottom). Apart from illuminating the object’s surroundings (the lamps, the lit house), emission also decorates (the glowing cracks, the blade, the tree’s lights) and informs (the screen’s content).
Figure 3: Overview of our data filtering pipeline. (1) Metadata filtering (1st row) keeps assets with an emissive texture or a non-zero emission factor (left) and drops the rest (right). (2) Heuristic filtering (2nd row) removes any asset whose emission map is entirely black (left) or identical to its albedo map (middle), or that duplicates another asset (right). (3) VLM filtering (3rd row) rejects non-emissive assets and assets with other defects, e.g., the multi-object asset is rejected. It agrees with human annotations on 87% of 404 held-out shapes.
Statistic
Value
# Shapes
36,826
# Categories
105
Train / Val / Test
36,426 / 200 / 200
Mean per-shape emissive ratio
8.1%
Median per-shape emissive ratio
1.4%
Table 1: Dataset statistics. The statistics of our curated dataset. Per-shape emissive ratio denotes the percentage of UV-occupied texels with non-zero emission.
Figure 4: Overview of the four emission-generation methods in our benchmark. Each method takes the non-emissive materials M , the geometry g , and a rendered image I , and generates the emission Le in its own representation: an emission map on the input UV atlas (top), per-voxel emission on the o-voxel grid (middle), or multi-view emission images that are back-projected onto the mesh (bottom). TRELLIS.2-Emission and SegviGen-Emission share the same pipeline and differ in the target: continuous emission e , or a mask M that is thresholded and multiplied by the albedo ( M⊙a ); their initialization and decoders are detailed in Secs. 4.3 and 4.5 .
Figure 5: User study. Each asset is rendered in a dark room with each baseline’s predicted emission (Pred 1 to Pred n ) ramping from 0 to 8. Participants rank the baselines by plausibility, with the ground-truth emission video and a well-lit rendering as reference (Ref).
Method
IoU ↑
MAE ↓
PSNR ↑
TEXGen-Emission
0.092
0.025
25.10
TRELLIS.2-Emission
0.271
0.021
26.81
Hunyuan3D-Emission
0.117
0.022
26.46
SegviGen-Emission
0.147
0.041
24.50
Table 2: Quantitative evaluation on the test set. We report the mean over five sampling seeds.
Figure 6: Qualitative comparison of the baselines. Each object is rendered in a dark room and lit only by its predicted emission; the leftmost column is the input image, and GT is rendered with ground truth emission. TRELLIS.2-Emission stays closest to the ground truth on all three assets, but still mislocates emissive surfaces or assigns inaccurate color (the ring’s emissive region is misplaced and it should glow yellow). The other baselines show their characteristic failures: TEXGen-Emission scatters emission across UV islands (spilled color on the helmet) and fails to keep non-emissive regions at exact zero (speckles on the helmet, faint glowing noise over the ring); Hunyuan3D-Emission shifts the emission color (the red beam turns white) and lights up regions that should stay dark (the saber handle and the ring band); and SegviGen-Emission over-extends or misses emissive regions (the ring’s whole band glows, while the helmet’s stripes stay unlit).
Method
Mean rank ↓
Top-1 rate ↑
TEXGen-Emission
3.22 ± 1.02
10.1%
TRELLIS.2-Emission
1.46 ± 0.76
67.2%
Hunyuan3D-Emission
2.50 ± 0.95
14.3%
SegviGen-Emission
2.82 ± 0.89
8.4%
Table 3: User study results. 14 participants each ranked the outputs of all four baselines on 53 objects. We report the mean rank and the fraction of rankings in which a method was placed first.
Appendix figures & tables16 assets
Supplementary material from the paper’s appendix.
Appendix
Figure S1 : VLM Prompt. The full prompt given to GPT-5.6 Luna [ 22 ] for every shape, adapted from the annotator guide.
Figure S2 : Examples rejected by the VLM. Each row from left to right shows the normal render (Object), the dark-box render with emission on, and the emit-only render.
Figure S3 : Word cloud for asset categories. The 105 categories of the final dataset. Larger words mark categories with more shapes. The other class (3.0%) is omitted.
Emission strength
Assets
Share
1 (default value)
66,693
80.2%
10
7,872
9.5%
other values
8,559
10.3%
total
83,124
100%
Appendix
Table S1 : Emission strength in the source assets. Per-asset maximum of the glTF emission strength over the 83k candidates. Most of them ( ∼ 80%) only have the default value of 1. ∼ 9.5% have a value of 10. The values for the remaining 10% are between 1 and 10.
Figure S4 : Emissive ratio before and after VLM filtering. Bars show the percentage of shapes within each set; the spike at 100% in the pre-VLM distribution is the fully emissive shapes, most of which the VLM stage removes.
VAE
MAE ↓
IoU ↑
PBR VAE
0.0461
0.694
Emission VAE
0.0330
0.746
Appendix
Table S2 : Emission VAE ablation. Reconstruction quality of the ground-truth emission with the pretrained PBR VAE and our emission VAE on the test set. MAE is measured on the ground-truth emissive voxels; IoU is computed over all occupied voxels.
Figure S5 : Sampled point cloud evaluation. We first sample 50,000 points from the original 3D asset. For the ground-truth emission, we read the color directly at the corresponding location on the original asset. For the predicted emission, we use nearest neighbor lookup. With the ground truth and the predicted emission, we apply a threshold to obtain the emission mask. MAE and PSNR compare the emission point clouds, and IoU compares the thresholded masks.
Figure S6 : User study interface. The top row is the reference: the object under standard lighting and the video with its ground-truth emission. The bottom row shows the four methods in a random order, labeled A–D. Participants rank them from 1 to 4 based on the plausibility of the emission. The participants can refer to the references as a guide.
Method
IoU ↑
Prec. ↑
Rec. ↑
F1 ↑
TEXGen-Emission
0.092
0.135
0.192
0.159
TRELLIS.2-Emission
0.271
0.361
0.487
0.415
Hunyuan3D-Emission
0.117
0.153
0.411
0.223
SegviGen-Emission
0.147
0.200
0.342
0.252
Appendix
Table S3 : Emissive-mask precision, recall, and F1. We report the mean over five random seeds of the generator. The emissive masks are obtained by a 1/255 threshold. TRELLIS.2-Emission is best on all four mask-based metrics, but no method localizes the emissive region well: even the best F1 is only 0.415.
Mask
All
Method
MAE ↓
PSNR ↑
MAE ↓
PSNR ↑
TEXGen-Emission
0.318
9.62
0.025
25.10
TRELLIS.2-Emission
0.264
11.07
0.021
26.81
Hunyuan3D-Emission
0.312
10.00
0.022
26.46
SegviGen-Emission
0.299
10.39
0.041
24.50
Appendix
Table S4 : MAE and PSNR on ground-truth emissive points (Mask) and on all points (All). We report the mean over five random seeds of the generator. A point is emissive if its ground-truth emission exceeds 1/255 . TRELLIS.2-Emission is slightly better, but no method matches the artist-authored emission well.
Figure S7 : Occlusion failure of the multi-view baseline. An emissive surface hidden in the rendered views is never observed by the model, so its correct emission can not be generated. Here, the bulb of the lamp is occluded; the predicted emission (top) is nearly dark, and the lamp fails to light the dark room in the rendering when we have the light on (bottom).
Figure S8 : Fracturing artifacts in the TEXGen-Emission model. Emission predicted in UV space leaks into nearby UV islands, which map back to surface regions that should not emit.
Figure S9 : Noise artifacts of Hunyuan3D-Emission. Compared with the dark room rendering of the ground truth (1st row left), the predicted hilt of the lightsaber glows white although it should not emit (1st row right). In one view of the multi-view emission output, the prediction (3rd row) differs from the ground truth (2nd row) only faintly, but the positive difference max(pred−GT,0) (4th row), amplified 10× (5th row), reveals non-zero emission over the whole non-emissive hilt.
Figure S10 : Category distribution. All 105 categories, sorted by size, on a log scale, split over two columns (the largest 53 on the left, the remaining 52 on the right) sharing one axis.
Figure S11 : Samples from the curated dataset. Thirty assets, one drawn at random from each of the thirty largest categories. Each cell shows the asset in our dark room setting, lit by its own emission; the card in the top right shows the same asset under standard lighting, and the line at the bottom left gives the category, its shape count, and its percentage of the dataset.
Figure S12 : Additional qualitative results. The leftmost column is the input image, GT is the rendering with the ground-truth emission, and each object is rendered in our dark room setting ( Appendix C ), lit by its predicted emission.
3D texture generation is receiving increasing attention, as it enables the creation of realistic and aesthetic texture materials for untextured 3D meshes. However, existing 3D texture generation methods are limited to producing only a few types of non-emissive PBR materials (e.g., albedo, metallic maps and roughness maps), making them difficult to replicate highly popular styles, such as cyberpunk, failing to achieve effects like realistic LED emissions. To address this limitation, we propose a novel task, emission texture generation, which enables the synthesized 3D objects to faithfully reproduce the emission materials from input reference images. Our key contributions include: first, We construct the Objaverse-Emission dataset, the first dataset that contains 40k 3D assets with high-quality emission materials. Second, we propose EmissionGen, a novel baseline for the emission texture generation task. Third, we define detailed evaluation metrics for the emission texture generation task. Our results demonstrate significant potential for future industrial applications. Dataset will be available at https://github.com/yx345kw/EmissionGen.
Zhiyuan Zhang, Zijian Zhou, Linjun Li +3
Peking University · XG Tech · King’s College London
High-fidelity image-to-3D generation requires a 3D representation that captures both geometry and appearance. However, preserving fine detail across the physically based rendering (PBR) modalities needed for relighting remains challenging. To address this, we propose Luce, a 3D representation that unifies geometry and PBR materials within a voxelized multimodal Gaussian cloud, using dedicated Gaussian primitives for albedo, metallic-roughness, and surface normals. A variational autoencoder compresses this representation into a unified material-aware latent space. A rectified-flow transformer generates this latent from a single image using multi-layer features from a pretrained image encoder that preserve both semantic context and fine spatial detail. The latent is then decoded into relightable PBR Gaussians and an optional textured mesh with a tangent-space normal map. On Toys4K, Luce achieves state-of-the-art single-image-to-3D generation, improving FID by 28% over the strongest baseline. We further evaluate Luce on a benchmark of AI-generated images depicting diverse subjects and materials, where it improves the CLIP image-alignment score over the best baseline (0.8519 vs. 0.8299). Luce generates relightable, geometrically accurate, and materially faithful assets that preserve fine details such as text, logos, and inscriptions.
Recent native 3D generators achieve high-quality asset creation by modeling structured 3D latents with flow-based transformers, but their high-resolution generation remains computationally expensive due to costly iterative sampling. In this work, we introduce a few-step distillation framework for native 3D generation, enabled by continuous self-consistency training with a custom JVP FlashAttention kernel and an analytic velocity regression objective for sparse structured 3D latents. To stabilize joint optimization, we further introduce an adaptive loss weighting strategy that automatically balances them without manual tuning. Beyond acceleration, we reveal a previously overlooked failure mode of few-step native 3D generation: sampling-induced latent distribution drift preserves geometric correctness but disrupts downstream PBR texture synthesis, causing texture artifacts on geometrically valid assets. To address this issue, we propose Channel Match, a training-free adaptive per-channel latent calibration method that restores the latent distribution without modifying either generator. We validate our approach on TRELLIS.2 at the 10243 resolution setting. Our two-step model achieves 4.5× faster geometry sampling while maintaining geometry quality comparable to the many-step teacher, and Channel Match further improves rendering quality by recovering the latent interface between geometry and appearance generation.