ECCV 2026

SuperVoxelGPT: Adaptive and Ordered 3D Tokenization
for Autoregressive Shape Generation

Yuan Li^1,2 Congyi Zhang^1† Xifeng Gao^2† Xiaohu Guo^1†

¹University of Texas at Dallas · ²LightSpeed († Corresponding authors)

Full narrated overview — motivation, method, and results (3:36).

SuperVoxelGPT teaser — **Adaptive supervoxel tokenization.** The shape itself decides where the tokens go — fine cells where the geometry is detailed, coarse cells where it is smooth.

TL;DR — Set-based tokenizers are compact but unordered; voxel grids are ordered but redundant. SuperVoxelGPT resolves the trade-off with a saliency-guided, deterministically ordered supervoxel partition, cutting sequence length to 12.8% of uniform voxel tokenization while matching state-of-the-art quality at full 1024³ — which is what makes high-resolution shape generation fast enough to be practical.

12.8%

of uniform voxel
sequence length

10×

average speedup
over prior methods

1,048

tokens per shape
at 1024³

4.6s

generation time
per shape

Abstract

Autoregressive multimodal large language models (MLLMs) enable 3D generation but struggle to scale to high-resolution shapes due to inadequate 3D tokenizations. Compact set-based representations discard deterministic spatial ordering, leading to ambiguous sequence prediction, while uniform or octree-based voxel grids preserve ordering at the cost of severe redundancy and excessively long sequences. This structural trade-off limits stable and efficient autoregressive 3D generation.

We present SuperVoxelGPT, a representation-first framework that resolves this tension through adaptive and deterministically ordered supervoxel tokenization. Given a prompt, we first predict a coarse geometric saliency distribution and construct a shape-adaptive supervoxel partition using saliency-guided centroidal Voronoi tessellation, allocating fine-grained cells to complex regions and larger cells to smooth regions. Conditioned on this prompt and ordered supervoxel layout, we introduce a SuperVoxelVAE and fine-tune a pretrained MLLM to autoregressively generate supervoxel tokens.

Experiments using Trellis-500K data show that SuperVoxelGPT reduces token sequence length to 12.8% of uniform voxel tokenization while achieving state-of-the-art generation quality and an average 10× speedup over prior methods.

Method

Inference is three steps. The key move is the order: the token budget is decided before the shape exists — saliency is predicted from the prompt alone, so the partition is available as a conditioning signal rather than something recovered after generation.

SuperVoxelGPT pipeline at inference — **Overall pipeline at inference.** The three numbered stages correspond to the steps below.

Inference procedure

Predict saliency. From the prompt alone, before any geometry exists — so the token budget is decided without first having to generate the whole shape.
Derive the partition by saliency-guided CVT. Centroidal Voronoi tessellation turns that field into supervoxels: dense cells on detailed regions, large cells on smooth ones. Being deterministic, it also fixes a single valid reading order.
Generate with the MLLM. Conditioned on the prompt and that ordered layout, a pretrained multimodal LLM emits one token per supervoxel, which the SuperVoxelVAE decodes back to geometry — and because the layout is known in advance, decoding runs in parallel.

Training procedure

Training has two parts. We first train the two VAEs — the Saliency VQ-VAE and the SuperVoxelVAE — so that each learns its own discrete vocabulary. We then freeze them and train the two generators on top: a MaskGIT that predicts the saliency tokens from the prompt, and a fine-tuned Qwen2.5-0.5B that emits the supervoxel tokens.

Two VQ-VAEs and their token correspondence — **The two tokenizers.** Each VAE learns a discrete vocabulary: the *Saliency Volume VAE* maps a 64³ volume to 8×8×8×2 = **1,024 tokens**, and the *SuperVoxel VAE* maps the shape to **one token per supervoxel**. Both quantise with FSQ.

Results

Text-to-3D generation

Method	CD_L2 ↓	PSNR ↑	ULIP-2 ↑	MCS ↑	Time (s) ↓	Resolution	Tokens ↓
BrickGPT	0.0851	12.99	0.199	15.39	215.5	20³	179
OctGPT	0.0797	19.80	0.329	11.74	165.8	256³	47,783
Ours	0.0134	32.05	0.469	44.19	4.48	1024³	1,065

Text-to-shape gallery — **Text-to-3D generation.** Compared with BrickGPT and OctGPT against ground truth. The two baselines fail in opposite ways, and the comparison isolates why: BrickGPT works at 20³ with only 179 tokens, so its outputs are recognisable but blocky — the budget is small because the *resolution* is small. OctGPT spends 47,783 tokens at 256³ and still smooths away thin structure, because a uniform-ordered grid spends most of its sequence on flat regions. Ours uses **1,065 tokens at 1024³** — fewer than either, at sixteen times OctGPT's resolution — because the partition concentrates cells on spires, rigging and ornament and leaves smooth panels coarse.

Image-to-3D generation

Method	CD_L2 ↓	PSNR ↑	ULIP-2 ↑	MCS ↑	Time (s) ↓	Resolution	Tokens ↓
CraftsMan3D	0.0252	22.90	0.448	37.85	7.92	512³	2,048
TRELLIS	0.0182	27.28	0.464	21.64	9.03	256³	8,147
Direct3D-S2	0.0168	26.14	0.461	33.17	137.20	1024³	55,420
TRELLIS2	0.0123	32.19	0.475	44.24	34.25	1024³	7,303
Ours	0.0122	32.08	0.474	44.21	4.60	1024³	1,048

At matched quality with the strongest baseline (TRELLIS2), our sequences are 7.0× shorter and generation is 7.4× faster, both at full 1024³ resolution.

Image-to-shape gallery — **Image-to-3D generation.** The same pipeline, conditioned on a single image instead of text. The interesting comparison here is with TRELLIS2, which reaches the same 1024³ and essentially the same quality (MCS 44.24 vs our 44.21) — the shapes are hard to tell apart, which is the point: we match it with **7.0× fewer tokens and 7.4× less time**. Direct3D-S2 shows the alternative cost of that resolution: 55,420 tokens and 137s per shape. Note also that saliency is predicted *from the image alone*, before any geometry exists, so the budget is allocated without first generating the whole shape.

Ablation

Two design choices are tested separately, each against a variant that removes it while holding everything else fixed.

Variant	CD_L2 ↓	PSNR ↑	ULIP-2 ↑	MCS ↑
Uniform CVT (Model 2 — same budget, no saliency guidance)	0.0157	29.09	0.469	40.13
CraftsMan3D + FSQ (Model 3 — set-based encoder instead of ours)	0.0402	20.91	0.364	21.73
Ours	0.0122	32.08	0.474	44.21

Both matter, but not equally. Dropping saliency guidance while keeping the same compression ratio (Model 2) costs 3.0 dB PSNR and 4.1 MCS — the budget is spent, just in the wrong places. Replacing our spatially localised encoding with a set-based one (Model 3) is far more damaging: 11.2 dB PSNR and half the MCS. Saliency guidance decides where tokens go; the supervoxel encoding is what makes the sequence learnable at all.

**Ablation.** Two baselines against ours, each removing one design choice. Comparing the partitions **(a)** and **(d)** shows what saliency guidance buys: at the identical compression ratio, (d) concentrates small cells where the geometry is complex while (a) spends the same budget evenly — and the generated result **(b)** loses exactly the detail that budget was spent away from. **(c)** isolates the tokenizer instead of the partition: our MLLM kept intact, but the SuperVoxelVAE swapped for a set-based encoder-decoder with the same FSQ layer.

Limitations

The method's central assumption — that a coarse saliency field predicted before generation is a good guide for where to spend tokens — is also where it breaks down. Two cases are worth stating plainly.

Predicted saliency — **(d)** Its saliency

Supervoxel partition — **(e)** Its partition

Detail below the saliency threshold is missed (a, b)

Saliency values under t = 0.1 are heavily coarsened, so structures that are genuinely fine but not salient are never flagged, receive only sparse cells, and get smoothed over. The ribbed tubing inside the box in (b) is gone in (a). The failure is quiet: the output is a clean, plausible shape that is simply missing something.

On noisy shapes the compression advantage disappears (c, d, e)

The compression comes from contrast in the saliency field, not from detail itself. On the densely perforated shape (c), the predicted saliency (d) is salient almost everywhere, so there is no contrast to exploit; the partition (e) degenerates towards a uniform one and the token budget approaches that of a plain voxel grid.

BibTeX

@inproceedings{li2026supervoxelgpt,
  title     = {SuperVoxelGPT: Adaptive and Ordered 3D Tokenization
               for Autoregressive Shape Generation},
  author    = {Li, Yuan and Zhang, Congyi and Gao, Xifeng and Guo, Xiaohu},
  booktitle = {Proceedings of the European Conference on Computer Vision (ECCV)},
  year      = {2026}
}