SILSA: Sliding-Window Slice Latents for Topology-Preserving High-Resolution 3D Generation

SILSA: Sliding-Window Slice Latents for Topology-Preserving High-Resolution 3D Generation

arXiv:2610.02201v1 [cs.CV] 1 Oct 2026

Tianjiao Yu, Xinzhuo Li, Yifan Shen, Ying Shen, Kiet A. Nguyen, Adheesh Sunil Juvekar, Ismini Lourentzou University of Illinois Urbana-Champaign {ty41,lourent2}@illinois.edu

Figure 1: High-resolution 3D generation through topology-preserving slice latents. Given a single input image, SILSA generates high-resolution 3D shapes with coherent structure and detailed local geometry, including thin structures, holes, repeated components, and long-range connectivity. By modeling shapes through compact overlapping slice latents and cross-axis volumetric coordination, SILSA maintains cross-sectional continuity and produces coherent global structure.

Abstract High-resolution 3D generation increasingly relies on voxel latents and multi-stage pipelines that first predict active structure and then synthesize local geometry. While effective, this design fragments continuous surfaces into many local tokens, inflates generation cost, and often weakens topological consistency for thin or highly connected shapes. We introduce SILSA, a topology-aware 3D generation framework that represents shapes with compact sliding-window slice latents. Instead of generating expensive voxel tokens, SILSA uses a fixed set of overlapping slices along the three canonical axes, where each token summarizes a local depth window to preserve cross-sectional continuity and support single-stage rectifiedflow generation. A Slice VAE encodes oriented surface samples into multi-axis slice latents and reconstructs them with a sparse volumetric decoder, while a Volumetric Anchor Lattice coordinates directional slice streams through a shared 3D workspace. To preserve structural correctness, we introduce slice-level topology supervision that matches persistence diagrams and aligns Betti transitions across neighboring slices. Experiments show that SILSA improves structural fidelity while substantially reducing generation cost. SILSA improves PSNR by 8.7%, coverage by 5.96 absolute points, and Betti error by 9.2% over the strongest baseline, while using 70.0% fewer tokens than the next-most compact baseline and over 98% fewer tokens than sparse or hierarchical tokenizers, effectively reducing training memory by 40.4% and inference time by 58.5%. Qualitative results further show improved preservation of thin structures, repeated components, and long-range connectivity. PLAN Lab https://plan-lab.github.io/silsa 40th Conference on Neural Information Processing Systems (NeurIPS 2026).

1

Introduction

High-resolution 3D generation has advanced rapidly from per-instance optimization toward generative models that synthesize complete 3D assets from a single image or text prompt. Early score-distillation and multi-view diffusion pipelines demonstrated that strong 2D generative priors can be lifted into plausible 3D objects [52, 38, 70, 43, 44, 45, 60]. More recent native 3D generators learn compact latent spaces and train diffusion, autoregressive, or flow models directly over 3D structure [33, 90, 85, 87, 55, 76, 25, 81]. These systems make 3D generation substantially faster and more scalable, but they still struggle to preserve fine 3D structure. Generated shapes often match the overall object appearance while breaking thin parts, openings, and repeated components. These failures alter connectivity, remove openings, and break part relationships. This is a representation challenge fundamental to high-resolution 3D generation. Dense voxel grids provide a direct spatial scaffold by representing shape as occupancy or signed-distance values on a regular 3D lattice [13, 74, 46], but their memory and computation grow cubically with resolution. Sparse voxel, octree, and hierarchical tokenizers reduce this cost by modeling only occupied regions, high-detail regions, or progressively refined geometry [41, 56, 67, 55, 76]. However, their token count can remain data-dependent and may grow for objects with thin supports, many repeated parts, or complex surface topology. Compact alternatives, including set-based shape latents [85, 90, 33], triplanes [5, 20, 24, 27], primitive-based representations [35, 66, 89] keep generation tractable, but they weaken the direct correspondence between a token and the local geometric structure it must preserve. As a result, a model may achieve low surface or rendering error while still producing a structurally incorrect shape, such as filling a hole, breaking a support, or merging two nearby components. To address this dilemma, we introduce SILSA, a 3D generation framework that represents highresolution shapes with sliding-window slice latents. The key insight is that a slice (with a small thickness) captures the topology of an entire planar region in one coherent unit, whereas existing representations either fragment this structure across many local cells or compress it into tokens with weakened spatial correspondence. A sequence of slices further preserves continuity along the slicing axis, directly exposing how connected components, holes, and part boundaries evolve along each canonical direction. Motivated by these properties, SILSA decomposes a shape into three sequences of overlapping slice latents along the x, y, and z axes. For N slice positions per axis, the generator operates on only 3N latent tokens, regardless of the object’s occupancy, surface area, or part complexity. Because each token summarizes a local depth window instead of an infinitesimal plane, the representation remains compact while still exposing thin parts, nearby surfaces, and small openings to the model. The three axis-wise sequences provide complementary cross-sectional views of the same object, giving SILSA a short, spatially indexed latent sequence for high-resolution 3D generation. To make this representation effective for generation, SILSA combines compact slice latents with crossaxis coordination and topology-aware training. First, a SliceVAE encodes oriented surface samples into overlapping slice latents and reconstructs them through a sparse volumetric decoder, preserving local surface geometry. A Volumetric Anchor Lattice then coordinates the three directional slice streams inside the rectified-flow transformer: each slice token reads from and writes to the anchor plane associated with its axis and depth position, allowing x-, y-, and z-aligned evidence to accumulate in a shared 3D workspace and form a single coherent shape. Finally, slice-level topology losses supervise the decoded cross-sections by matching persistent-homology structure and aligning Betti transitions across adjacent slices [19, 92, 30, 16, 65, 64]. Together, these components allow SILSA to preserve both geometric fidelity, such as accurate surfaces and part shapes, and topological structure, such as connected components, holes, and consistent connectivity across depth. Experiments show that SILSA improves high-resolution image-to-3D generation while substantially reducing generation cost. Across both settings, SILSA preserves fine structures that are commonly degraded by compact 3D latents, including thin supports, handles, holes, spokes, railings, and repeated parts. Empirically, SILSA improves both structural fidelity and efficiency. On image-conditioned 3D generation, it achieves the best FD, PSNR, coverage, and MMD, with an 8.7% relative gain in PSNR and a 5.96-point absolute gain in coverage over the strongest baselines, while matching the best KD and LPIPS. The SliceVAE further reduces Betti error by 9.2% relative to the strongest reconstruction baseline, indicating better preservation of connected components and holes. At the same time, SILSA uses only 384 fixed slice tokens, 70.0% fewer than the next-most compact baseline, reducing training memory by 40.4% and inference time by 58.5%. Qualitative results further show that cross2

axis coordination through the Volumetric Anchor Lattice reduces inconsistent slice predictions and produces more coherent 3D assets. In summary, the contributions of our work are: • We introduce SILSA, a topology-aware image-to-3D generation framework that encodes shapes into a compact set of spatially grounded sliding-window slice latents across three canonical axes. • We design a topology-aware SliceVAE that combines overlapping slice aggregation, sparse volumetric decoding, persistent-homology matching, and Betti-transition supervision to preserve surface geometry, connected components, and hole structures. • We develop a single-stage rectified-flow generator with a Volumetric Anchor Lattice, enabling cross-axis coordination through shared spatial memory and efficient image-conditioned 3D generation from only 384 slice tokens.

2

Related Work

High-resolution 3D generation has evolved from lifting 2D diffusion priors through score distillation and differentiable rendering [52, 38, 70, 7] to image-conditioned multi-view reconstruction pipelines [44, 45, 60, 42, 61] and native 3D generators over learned latent spaces [13, 33, 90, 87, 76, 75, 25, 80]. Existing 3D representations trade off efficiency and structure: dense voxels provide spatial grounding but scale cubically [74, 13], triplanes and set latents improve compactness but weaken local geometric correspondence [5, 20, 85], and sparse or hierarchical voxel tokenizers preserve locality but require data-dependent token counts and often multi-stage generation [55, 76, 25]. Crosssectional representations provide a spatially grounded alternative, as planar slices expose components, holes, and connectivity changes that compact global latents can blur, while OReX [57] demonstrates that such slices provide useful geometric cues for reconstruction. SILSA differs by using multi-axis cross-sections as a learned generative latent with fixed sliding-window slice tokens. Moreover, our topology supervision builds on persistent homology and Betti-based losses for preserving connectivity and holes [19, 92, 30, 16, 64], but avoids expensive full-volume topology matching by supervising persistence within slices and Betti transitions across neighboring slices. Additional discussion of 3D generation, latent 3D representations, and topology-aware learning is provided in Appendix A.

3

Method

State-of-the-art 3D generative models encode shapes into structured latent tokens and generate them with transformer-based diffusion or rectified-flow models [76, 25]. The token count, however, scales with surface area, making generation expensive and requiring multi-stage pipelines that first predict which voxels are active before generating their structured latents. Beyond efficiency, voxellevel tokenization also fragments continuous surfaces into many local elements, making topological coherence challenging to model. We propose SILSA to address these limitations by generating compact, spatially grounded slice latents (Figure 2). First, we introduce a SliceVAE that maps 3D shapes to compact multi-axis slice latents and decodes them into a high-resolution mesh (§3.1). The encoder aggregates surface points with overlapping sliding windows along the three canonical axes, yielding a compact set of tokens that preserves local cross-sectional structure. The decoder populates a volumetric feature grid from these tokens and reconstructs geometry through sparse volumetric upsampling. To preserve structural correctness, we further introduce a Slice-Wise Topology-Preserving Loss that supervises decoded cross-sections (§3.2). Second, we train a rectified flow transformer to generate slice latents from an input image (§3.3). Since the slice layout is fixed, generation does not require a separate active-voxel prediction stage. Instead, we introduce a Volumetric Anchor Lattice (VAL), a shared spatial memory that enables slice tokens from different axes to read and write axis-aligned anchor planes during denoising. 3.1

Topology-Aware Slice VAE

Sliding-Window Slice Encoder. Inspired by previous works [25, 58], which aggregate point cloud features into sparse voxels via PointNet [53], we adopt the same local pooling paradigm for encoding 3D geometry. However, we replace structured voxels with N axis-aligned slices (planar bins) along each canonical axis. Each slice token summarizes the local geometry within a depth interval, reducing the representation to 3N tokens in total (N tokens for each of the x, y, and z axes). The three axis-wise slice sequences provide complementary geometric evidence that the decoder fuses for 3

3D Mesh

Window X-2

X-axis Slices

Window X-1

Reconstructed HighQuality Mesh

Window X

Image Encoder

Image Embedding ...

Conditioning (2D image) ... Window X-1

Rectified Flow Transformer t=T (Noise)

Window X

VAE Encode

...

...

...

Transformer Refinement & Sparse Upsampling

Scatter & Fuse

...

...

Z-axis Slices

t=T (Data)

Transformer Block

Window X-2

Y-axis Slices

Transformer Block

...

Dual Marching Cubes

...

Window X-2

Window X-1

Volumetric Anchor Lattice (VAL)

Window X

Read Process

...

Write

...

16*16*16 Grid

256*256*256 Grid

Slice-wise Topology Loss

16*16*16 Grid

Token Aggregation

(i) Sliding-window Slicing and VAE Encoder

(ii) Flow-based Generation

(iii) Reconstruction

Figure 2: SILSA Overview. SILSA represents each shape with overlapping slice latents along the x, y, and z axes. A topology-aware SliceVAE encodes surfaces into fixed multi-axis slice latents and decodes them through sparse volumetric upsampling into a high-resolution mesh. An imageconditioned rectified-flow transformer generates these latents, using a Volumetric Anchor Lattice as shared 3D memory for cross-axis coordination. Slice-wise persistent-homology and Betti-transition losses supervise connected components, holes, and topological consistency during VAE training. faithful reconstruction. A single bin, however, may contain very few points. To provide sufficient geometric context, we encode each bin using a sliding window of w surrounding bins. N

N

p p Formally, given a 3D mesh, we sample a point cloud P = {pℓ }ℓ=1 with normals {nℓ }ℓ=1 and partition the bounding box into N = 128 bins per axis. For bin k along axis j, the window gathers all points within w/2 bins on either side: o n (1) Pkj = p ∈ P k − w2 ≤ ⌊p(j) · N ⌋ < k + w2 , k = 0, . . . , N −1,

with boundary bins clamped to [0, N −1]. Each point p ∈ Pkj is augmented with its depth-relative offset δ(p) = ⌊p(j) · N ⌋ − k, which indicates its displacement from the center bin. A shared MLP processes each augmented point independently, and the window representation is obtained by max-pooling the resulting point features: hjk = maxj MLPϕ (p, np , δ(p)) ∈ Rd .

(2)

p∈Pk

The pooled feature hjk is mapped to posterior parameters (µjk , log σ jk ), defining a Gaussian slice latent    qϕ (zjk | P) = N µjk , diag (σ jk )2 . (3) During training, the decoder receives latent samples zjk ∼ qϕ (zjk | P). After training, we use the posterior mean µjk as the deterministic slice latent for flow training. With w = 8, each bin feature is informed by points spanning 8 consecutive slices. The full latent representation is Z = {zjk }, yielding 3N = 384 slice latents. For bins whose entire window is empty, we assign a learned empty embedding. Decoder. To reconstruct geometry from the slice latents, we first scatter the three axis-wise latent sequences into a shared coarse 3D feature grid. Because the slice resolution can be higher than the grid resolution, multiple neighboring slice latents are mapped to the same coarse grid plane. Let Ku = {k | ⌊kD/N ⌋ = u} denote the set of slice indices mapped to grid plane u. We aggregate the projected slice latents by normalized summation: 1 X Πa (u) += Wa zak , (4) |Ku | k∈Ku

where a ∈ {x, y, z} denotes the slice axis, and Πa (u) denotes the corresponding grid plane, i.e., Πx (u) = G[u, :, :], Πy (u) = G[:, u, :], and Πz (u) = G[:, :, u]. This scatter operation fuses the three axis-wise slice decompositions into a shared volumetric representation. A sparse transformer decoder then refines these features, followed by two self-pruning 4

upsampling stages [55] that progressively subdivide active cells and prune empty regions, increasing the grid resolution from 163 to 2563 . At the final resolution, a linear head predicts per-cell isosurface parameters, including SDF values, vertex deformations, and interpolation weights. The output mesh is then extracted via differentiable Dual Marching Cubes [58, 35]. VAE training. The SliceVAE is trained end-to-end with differentiable rendering losses Lrender = λd Ld + λn Ln + λm Lm , where Ld , Ln , and Lm are L1 losses on depth, normal, and silhouette maps, respectively. The latent space is regularized by a KL divergence term: LKL =

−1 X NX j

  DKL qϕ (zjk | P) ∥ N (0, I) .

(5)

k=0

Additionally, we propose a slice-wise topology-preserving loss Ltopo (§3.2) that supervises the topological correctness of decoded cross-sections. The full training objective is LVAE = Lrender + βKL LKL + λtopo Ltopo . 3.2

(6)

Slice-Wise Topology-Preserving Loss

Standard rendering losses capture local surface discrepancies but are often insensitive to structural failures in thin or highly connected shapes, such as bicycle wheels with dense spokes or plants with many branching stems. We therefore introduce a slice-wise topology-preserving loss that supervises the topology of decoded cross-sections during VAE training. Specifically, we use persistent homology [92, 19] to compare per-slice persistence diagrams and align transitions across neighboring slices. We provide a visual illustration of the multi-axis topology signals used by our loss in Appendix B. After upsampling, the decoder predicts SDF values on a dense corner grid. For efficiency, we compute the topology loss on Ns evenly spaced cross-sections along each canonical axis. For a sampled depth index k, we extract a cross-section by indexing the SDF grid and converting to a soft occupancy map ŜkX (y, z) = σ(−ŝ[k, y, z]/κ), with analogous definitions for ŜkY and ŜkZ . Here, ŝ is the predicted SDF grid and κ > 0 controls the sharpness of the occupancy boundary. Ground-truth cross-sections are obtained by evaluating signed distances from the target mesh on the same grid and applying the same SDF-to-occupancy conversion. Topological Signature. Each cross-section Ŝkj induces a superlevel-set filtration, whose persistence diagram Dgmd (Ŝkj ) records d-dimensional topological features as birth–death pairs (bp , dp ), where d=0 corresponds to connected components and d=1 corresponds to holes. The persistence of a feature is pers(p) = bp − dp under the superlevel convention. In addition to per-slice persistence diagrams, we compute the Betti number at the occupancy boundary,
j Bk,d = βd {Ŝkj ≥ 0.5} , (7) j j j and define the transition sequence ∆Bk,d = Bk+1,d − Bk,d , which captures where cross-sectional topology changes along axis j, i.e., where connected components or holes appear, disappear, merge, or split as the slicing plane moves through the shape. By Morse theory, nonzero transitions correspond to intervals containing critical events of the height function along axis j [48]. We supervise both the per-slice persistence diagrams and the Betti transition sequences against the corresponding ground-truth cross-sections.

Topological Losses. We use two complementary losses to supervise the topology of decoded cross-sections. The per-slice topology matching term preserves the topology within each decoded cross-section by matching predicted and ground-truth persistence diagrams: LPH =

s −1 X NX X

j

k=0 d∈{0,1}

min γ∈Γ

X

∥p − γ(p)∥22 ,

(8)

p∈Dgmd (Ŝkj )

where j indexes the slicing axis, k indexes the sampled cross-section, and d denotes the homology dimension, with d=0 for connected components and d=1 for holes. The matching set Γ includes assignments to ground-truth topological features as well as to the diagonal, so unmatched predicted features are penalized according to their persistence. This suppresses spurious short-lived components and holes while preserving persistent structures that define the slice topology. 5

The inter-slice transition matching term preserves how topology evolves as the slicing plane moves through the shape. While per-slice matching encourages each decoded cross-section to have the correct connected components and holes, it does not explicitly enforce where these structures appear, disappear, split, or merge along the depth axis. We therefore supervise the Betti transition sequence: Ltrans =

s −2 X NX X 

j

2 j j ∆Bk,d (Ŝ) − ∆Bk,d (S) .

(9)

k=0 d∈{0,1}

j j j Here, ∆Bk,d = Bk+1,d − Bk,d records the change in the d-dimensional Betti number between adjacent slices along axis j. Matching these transitions encourages topological events to occur at the correct depths, reducing errors such as holes closing too early, thin supports disconnecting, or nearby j parts merging into spurious bridges. Since Betti counts are discrete, we compute Bk,d from the thresholded occupancy in the forward pass and use a straight-through estimator during backpropagation [2].

The final topology objective combines the two complementary terms: Ltopo = λPH LPH + λtrans Ltrans .

(10)

The persistence term LPH preserves the topology of individual cross-sections by matching connected components and holes in persistence-diagram space, while the transition term Ltrans preserves where these structures appear, disappear, split, or merge across neighboring slices. Together, they encourage the decoded geometry to match both the local topology of each slice and the global evolution of topology along each canonical axis. 3.3

Rectified Flow Generation with Volumetric Anchors

With the SliceVAE trained, we freeze the encoder and decoder and train a rectified flow transformer to generate slice latents from a single input image. The ground-truth latents Z (0) are obtained by encoding each training shape with the frozen encoder. The transformer learns to map noise to these latents, conditioned on image features. Following rectified flow [39]:   2 min Et,ϵ,Z (0) vθ Z (t) , t, cimg − (ϵ − Z (0) ) , (11) θ

where Z

(t)

= (1 − t)Z

(0)

2

+ tϵ and cimg is extracted by a frozen DINOv2 encoder [50].

Volumetric Anchor Lattice (VAL). The central challenge in generating multi-axis slice latents is cross-axis consistency: the three axis decompositions must describe one coherent 3D shape, yet each axis-wise sequence is denoised as a separate ordered set of slice tokens. We address this challenge with a Volumetric Anchor Lattice (VAL), a persistent 3D feature grid G ∈ RD×D×D×C that serves as shared spatial memory inside the rectified-flow transformer. At each transformer block, slice tokens read from and write to axis-aligned planes in this grid according to their slice axis and depth position. Tokens from different axes therefore deposit evidence into the same spatial workspace, and later blocks can retrieve this accumulated evidence to coordinate denoising across axes. This turns cross-axis consistency into a spatially grounded communication mechanism, encouraging the generated x-, y-, and z-aligned slices to decode into one coherent 3D object. Each slice token has a well-defined spatial footprint in the grid: an x-slice at depth k maps to the plane G[k ′ , :, :], a y-slice to G[:, k ′ , :], a z-slice to G[:, :, k ′ ], where k ′ = ⌊kD/N ⌋. We denote this depth plane as Gj (k) for axis j. This correspondence is geometric and requires no learning. Transformer block. Each block executes five operations. (1) Intra-axis Self-attention: for each axis independently, the N tokens attend to each other. (2) Anchor Read: each token cross-attends to its depth plane Gj (k) in the VAL, retrieving D2 anchor features that encode what other axes have written to the same spatial region. (3) Anchor Write: each token updates its depth plane via a gated mechanism: Gj (k) ← (1 − zg ) ⊙ Gj (k) + zg ⊙ fw (h̃jk , gkj ),

zg = σ(Wg [h̃jk ; gkj ]),

(12)

where h̃jk is the token after self-attention, gkj is the retrieved anchor feature, fw (h̃jk , gkj ) ∈ RC is broadcast to all cells in the corresponding depth plane, and zg ∈ RC is applied channel-wise. (4) Image Cross-attention to cimg . (5) Feed-forward Network. We stack L blocks; the VAL is initialized to zeros at each velocity evaluation and accumulates cross-axis evidence across transformer blocks. 6

Table 1: Image-to-3D generation. We compare SILSA with representative image-conditioned and native 3D generation methods. Best and second best highlighted.

Model

CLIP ↑

FD ↓

KD ↓

PSNR ↑

LPIPS ↓

COV(%) ↑

MMD(‰) ↓

Shap-E LN3Diff Direct3D 3DTopia-XL InstantMesh Gau.Any. XCube Dora SAR3D Trellis SparseFlex

80.16 82.79 74.12 76.46 84.41 80.91 84.91 80.35 84.67 85.03 88.22

34.64 26.98 24.97 24.21 20.13 22.46 10.32 22.84 22.12 10.31 11.16

0.87 0.76 0.33 0.29 0.29 0.44 0.09 0.23 0.18 0.08 0.08

16.84 18.73 22.36 22.06 25.72 23.84 23.99 24.68 26.31 24.01 30.12

0.21 0.19 0.17 0.18 0.11 0.15 0.13 0.13 0.10 0.14 0.05

61.41 55.21 58.72 58.93 66.84 60.01 73.01 67.42 70.30 72.10 73.12

19.19 19.84 18.46 17.62 16.72 15.47 14.92 15.63 15.12 14.36 14.52

SILSA

87.94

10.16

0.08

32.74

0.05

79.08

14.02

4

Experiments

Experiment Setup. We use Trellis-500K [76] for training. For evaluation, we use 200 randomly sampled Toys4K assets[63] and 50 in-the-wild images, with no overlap with the training set. We compare the reconstruction quality of Slice VAE with Dora [8], XCube [55], Trellis [76], and SparseFlex [25]. We use Chamfer Distance (CD) and F-score with thresholds of 0.01 and 0.005, and Betti Error [64, 30] to assess geometric fidelity and topological correctness, respectively. For image-to-3D generation, we compare with representative open-source methods, including ShapeE [33], LN3Diff [36], Direct3D [73], 3DTopia-XL [10], InstantMesh [77], GaussianAnything [84], XCube [55], Dora [8], SAR3D [9], Trellis [76], SparseFlex [25]. We use CLIP similarity [54] to evaluate input-output alignment. Overall generative quality is measured with FD [26] and KD [3], while PSNR and LPIPS capture reconstruction-level visual fidelity. We further report COV and MMD [1] to assess distributional fidelity. Full implementation details are provided in Appendix C. Image-to-3D. Table 1 shows that SILSA achieves the strongest overall image-to-3D generation performance while using a compact fixed-length slice representation. Compared with prior imageconditioned and native 3D generators, SILSA obtains the best FD, PSNR, coverage, and MMD, while matching the best KD and LPIPS. In particular, SILSA improves PSNR from 30.12 to 32.74 over SparseFlex, an 8.7% relative gain, and increases coverage from 73.12% to 79.08%, a 5.96-point absolute improvement. These gains indicate that the generated shapes are not only closer to the target distribution, but also preserve higher-fidelity geometry and broader structural diversity. Although SparseFlex attains a slightly higher CLIP score, SILSA achieves substantially better geometric and distributional metrics, suggesting that the proposed slice-latent representation improves 3D fidelity without sacrificing image alignment. VAE Reconstruction Evaluation. Table 2 evaluates the reconstruction quality of the proposed SliceVAE using geometric, volumetric, and topology-aware metrics. We report Chamfer Distance (CD), F-Score at thresholds τ =0.01 and τ =0.005, IoU, and Betti-Err, where Betti-Err is computed as the average test-set mismatch in connected components and holes. SILSA consistently outperforms strong 3D tokenizers across all metrics, achieving the lowest CD, highest F-Scores, highest IoU, and lowest Betti-Err. Compared with SparseFlex, the strongest baseline, SILSA reduces CD from 0.61 to 0.59, improves F-Score@0.01 from 96.18 to 96.79, improves F-Score@0.005 from 83.62 to 84.03, and increases IoU from 92.54 to 93.01. More importantly, Betti-Err decreases from 1.743 to 1.582, a 9.2% relative reduction, showing that SliceVAE better preserves connected components and holes rather than only improving surface-level reconstruction fidelity. Efficiency analysis. Table 3 highlights the efficiency advantage of generating fixed slice latents instead of variable active voxel tokens. SILSA uses only 384 tokens, which is over 98% fewer than Trellis, XCube, and SparseFlex. Despite this large reduction in sequence length, SILSA achieves the best reconstruction quality. The compact representation also translates directly into lower computational cost: relative to Dora, SILSA reduces training memory from 14.6GB to 8.7GB, training 7

Table 2: VAE reconstruction quality. We compare SILSA against representative 3D reconstruction tokenizers using geometric, volumetric, and topology-aware metrics. Betti-Err measures the average mismatch in connected components and holes. Best and second best results highlighted.

Model

CD ↓

F-Score@0.01 ↑

F-Score@0.005 ↑

IoU ↑

Betti-Err ↓

Dora XCube Trellis (SLAT) SparseFlex

9.76 3.21 1.12 0.61

67.92 81.74 92.83 96.18

38.71 52.06 71.45 83.62

64.85 78.13 86.29 92.54

4.916 4.382 2.871 1.743

SILSA

0.59

96.79

84.03

93.01

1.582

Table 3: Efficiency comparison. Token counts for variable-length methods are reported as mean ± std over the test set. Memory and training time are measured with batch size 4 on a single A100. Inference time is reported end-to-end per shape. Best and second best highlighted. #Tokens ↓

Stages ↓

Params (M) ↓

Train Mem. ↓ (GB)

Train Time ↓ (s/iter)

Inference ↓ (s/shape)

CD ↓

Dora XCube Trellis SparseFlex

1,280 64,821 19,847 ± 4,312 87,453 ± 18,264

1 2 2 2

124 87 347 213

14.6 68.4 42.7 55.4

0.38 1.18 0.71 1.47

0.82 2.74 1.93 1.15

9.76 3.21 1.12 0.61

SILSA

384 (N =128)

1

96

8.7

0.21

0.34

0.59

Model

time from 0.38s/iter to 0.21s/iter, and inference time from 0.82s/shape to 0.34s/shape. Additionally, SILSA uses a one-stage generator, making high-resolution generation faster and more predictable. Qualitative Results. Qualitatively, SILSA better preserves the structural details that are most easily lost in compact 3D latent spaces. For image-to3D generation, Figure 3 shows that SILSA recovers plausible and viewconsistent 3D structure from a single image across diverse in-the-wild examples. These results indicate that multi-axis sliding-window slice latents preserve local cross-sectional Figure 3: Image-to-3D generation in the wild. structure while maintaining global coherence. For VAE reconstruction, Figure 4 shows that SILSA stays closer to the ground-truth geometry on objects with thin supports, articulated parts, dense branches, holes, and repeated structures, while competing tokenizers often smooth fine details, merge nearby components, or distort fragile parts. Additional examples can be found in Appendix D. *7

Ablations. Table 4 validates the im- Table 4: Key ablations. We ablate topology supervision portance of spatially grounded cross- and VAL. Best and second best highlighted. axis communication. When the three Variant CD ↓ IoU ↑ Betti ↓ slice streams are generated independently, performance degrades substanSlice-wise topology loss No Ltopo 0.78 81.76 4.43 tially, with CD increasing to 5.87, IoU Only LPH 0.72 84.41 2.91 dropping to 49.26, and Betti-Err risOnly Ltrans 0.69 87.16 2.87 ing to 17.91. Direct cross-attention beFull Ltopo 0.59 93.01 1.58 tween axes improves consistency, reducCross-axis communication ing CD to 0.60 and Betti-Err to 1.64 No communication 5.87 49.26 17.91 while increasing IoU to 93.18. However, Direct cross-attn. 0.60 93.18 1.64 VAL achieves the best geometric and VAL 0.59 93.01 1.58 topological fidelity, obtaining the lowVAL resolution est CD of 0.59 and lowest Betti-Err of D=8 0.65 91.37 1.86 1.58, showing that a shared volumetric D=16 (default) 0.59 93.01 1.58 workspace provides more reliable coD=32 0.60 93.18 1.62 ordination than unconstrained token-to8

GT

SparseFlex

Trellis

Xcube

Dora

Figure 4: VAE reconstruction quality. We compare reconstructed meshes from SILSA and representative 3D models. Normal maps are shown in the top-right inset, and surface-error maps are shown in the bottom-right inset. Surface error is visualized as from low (blue) to high (red).

token attention. The VAL resolution ablation further shows that D=16 gives the best trade-off: compared with D=8, it reduces CD from 0.65 to 0.59, improves IoU from 91.37 to 93.01, and lowers Betti-Err from 1.86 to 1.58. Increasing the resolution to D=32 slightly improves IoU to 93.18, but worsens CD to 0.60 and Betti-Err to 1.62, suggesting that finer anchors add communication cost without improving structural correctness. The full topology loss gives the best reconstruction trade-off, reducing Betti-Err from 4.43 to 1.58 relative to removing Ltopo while also improving CD from 0.78 to 0.59. Additional ablations in Appendix E.

5

Conclusion

We introduce SILSA, a topology-aware image-to-3D framework for high-resolution image-to-3D generation that represents shapes with compact sliding-window slice latents along the three canonical axes. A SliceVAE reconstructs high-resolution geometry with slice-wise persistence and Bettitransition supervision, while a Volumetric Anchor Lattice coordinates directional slice streams inside a single-stage rectified-flow generator. Across reconstruction, generation, and efficiency evaluations, SILSA improves geometric fidelity, preserves thin and highly connected structures, and substantially reduces token count, memory, and inference cost. These results show that cross-sectional slice latents offer a compact and structurally faithful representation for scalable 3D generation. 9

References [1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In International conference on machine learning, pages 40–49. PMLR, 2018. [2] Yoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013. [3] Mikołaj Bińkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. arXiv preprint arXiv:1801.01401, 2018. [4] Nick Byrne, James R Clough, Israel Valverde, Giovanni Montana, and Andrew P King. A persistent homology-based topological loss for cnn-based multiclass segmentation of cmr. IEEE Transactions on Medical imaging, 42(1):3–14, 2023. [5] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometryaware 3d generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16123–16133, 2022. [6] Hansheng Chen, Jiatao Gu, Anpei Chen, Wei Tian, Zhuowen Tu, Lingjie Liu, and Hao Su. Single-stage diffusion nerf: A unified approach to 3d generation and reconstruction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2416–2425, 2023. [7] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 22246–22256, 2023. [8] Rui Chen, Jianfeng Zhang, Yixun Liang, Guan Luo, Weiyu Li, Jiarui Liu, Xiu Li, Xiaoxiao Long, Jiashi Feng, and Ping Tan. Dora: Sampling and benchmarking for 3d shape variational auto-encoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16251–16261, 2025. [9] Yongwei Chen, Yushi Lan, Shangchen Zhou, Tengfei Wang, and Xingang Pan. Sar3d: Autoregressive 3d object generation and understanding via multi-scale 3d vqvae. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28371–28382, 2025. [10] Zhaoxi Chen, Jiaxiang Tang, Yuhao Dong, Ziang Cao, Fangzhou Hong, Yushi Lan, Tengfei Wang, Haozhe Xie, Tong Wu, Shunsuke Saito, et al. 3dtopia-xl: Scaling high-quality 3d asset generation via primitive diffusion. pages 26576–26586, 2025. [11] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5939–5948, 2019. [12] Zilong Chen, Yikai Wang, Feng Wang, Zhengyi Wang, and Huaping Liu. V3d: Video diffusion models are effective 3d generators. arXiv preprint arXiv:2403.06738, 2024. [13] Yen-Chi Cheng, Hsin-Ying Lee, Sergey Tulyakov, Alexander G Schwing, and Liang-Yan Gui. Sdfusion: Multimodal 3d shape completion, reconstruction, and generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4456–4465, 2023. [14] Christopher B Choy, Danfei Xu, JunYoung Gwak, Kevin Chen, and Silvio Savarese. 3d-r2n2: A unified approach for single and multi-view 3d object reconstruction. In European conference on computer vision, pages 628–644. Springer, 2016. [15] Asger Nyman Christiansen, J Andreas Bærentzen, Morten Nobel-Jørgensen, Niels Aage, and Ole Sigmund. Combined shape and topology optimization of 3d structures. Computers & Graphics, 46:25–35, 2015. 10

[16] James R Clough, Nicholas Byrne, Ilkay Oksuz, Veronika A Zimmer, Julia A Schnabel, and Andrew P King. A topological loss function for deep-learning based image segmentation using persistent homology. IEEE transactions on pattern analysis and machine intelligence, 44(12):8766–8778, 2022. [17] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksen, Himanshu Arora, et al. Abo: Dataset and benchmarks for real-world 3d object understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21126–21136, 2022. [18] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objaverse-xl: A universe of 10m+ 3d objects. Advances in Neural Information Processing Systems, 36:35799–35813, 2023. [19] Edelsbrunner, Letscher, and Zomorodian. Topological persistence and simplification. Discrete & computational geometry, 28(4):511–533, 2002. [20] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12479–12488, 2023. [21] Huan Fu, Rongfei Jia, Lin Gao, Mingming Gong, Binqiang Zhao, Steve Maybank, and Dacheng Tao. 3d-future: 3d furniture shape with texture. International Journal of Computer Vision, 129(12):3313–3337, 2021. [22] Rickard Brüel Gabrielsson, Bradley J Nelson, Anjan Dwaraknath, and Primoz Skraba. A topology layer for machine learning. In International Conference on Artificial Intelligence and Statistics, pages 1553–1563. PMLR, 2020. [23] Zhirui Gao, Renjiao Yi, Yuhang Huang, Wei Chen, Chenyang Zhu, and Kai Xu. Partgs: Learning part-aware 3d representations by fusing 2d gaussians and superquadrics. arXiv preprint arXiv:2408.10789, 2024. [24] Anchit Gupta, Wenhan Xiong, Yixin Nie, Ian Jones, and Barlas Oğuz. 3dgen: Triplane latent diffusion for textured mesh generation. arXiv preprint arXiv:2303.05371, 2023. [25] Xianglong He, Zi-Xin Zou, Chia-Hao Chen, Yuan-Chen Guo, Ding Liang, Chun Yuan, Wanli Ouyang, Yan-Pei Cao, and Yangguang Li. Sparseflex: High-resolution and arbitrary-topology 3d shape modeling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14822–14833, 2025. [26] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. [27] Yicong Hong, Kai Zhang, Jiuxiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400, 2023. [28] Jiangbei Hu, Ben Fei, Baixin Xu, Fei Hou, Weidong Yang, Shengfa Wang, Na Lei, Chen Qian, and Ying He. Topology-aware latent diffusion for 3d shape generation. arXiv preprint arXiv:2401.17603, 2024. [29] X Hu, Y Wang, F Li, D Samaras, and C Chen. Topology-aware segmentation using discrete morse theory. In International Conference on Learning Representations (ICLR), 2021. [30] Xiaoling Hu, Fuxin Li, Dimitris Samaras, and Chao Chen. Topology-preserving deep image segmentation. Advances in neural information processing systems, 32, 2019. [31] Zixuan Huang, Mark Boss, Aaryaman Vasishta, James M Rehg, and Varun Jampani. Spar3d: Stable point-aware reconstruction of 3d objects from single images. arXiv preprint arXiv:2501.04689, 2025. 11

[32] Anushrut Jignasu, Ethan Herron, Zhanhong Jiang, Soumik Sarkar, Chinmay Hegde, Baskar Ganapathysubramanian, Aditya Balu, and Adarsh Krishnamurthy. Stitch: Surface reconstruction using implicit neural representations with topology constraints and persistent homology. arXiv preprint arXiv:2412.18696, 2024. [33] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023. [34] Mukul Khanna, Yongsen Mao, Hanxiao Jiang, Sanjay Haresh, Brennan Shacklett, Dhruv Batra, Alexander Clegg, Eric Undersander, Angel X Chang, and Manolis Savva. Habitat synthetic scenes dataset (hssd-200): An analysis of 3d scene scale and realism tradeoffs for objectgoal navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16384–16393, 2024. [35] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics (ToG), 39(6):1–14, 2020. [36] Yushi Lan, Fangzhou Hong, Shuai Yang, Shangchen Zhou, Xuyi Meng, Bo Dai, Xingang Pan, and Chen Change Loy. ln3diff: Scalable latent neural fields diffusion for speedy 3d generation. In European Conference on Computer Vision, pages 112–130, 2024. [37] Weiyu Li, Jiarui Liu, Rui Chen, Yixun Liang, Xuelin Chen, Ping Tan, and Xiaoxiao Long. Craftsman: High-fidelity mesh generation with 3d native generation and interactive geometry refiner. arXiv preprint arXiv:2405.14979, 2024. [38] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 300–309, 2023. [39] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747, 2022. [40] Fangfu Liu, Wenqiang Sun, Hanyang Wang, Yikai Wang, Haowen Sun, Junliang Ye, Jun Zhang, and Yueqi Duan. Reconx: Reconstruct any scene from sparse views with video diffusion model. arXiv preprint arXiv:2408.16767, 2024. [41] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems, 33:15651–15663, 2020. [42] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Mukund Varma T, Zexiang Xu, and Hao Su. One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. Advances in Neural Information Processing Systems, 36:22226–22246, 2023. [43] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9298–9309, 2023. [44] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023. [45] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9970–9980, 2024. [46] Nissim Maruani, Wang Yifan, Matthew Fisher, Pierre Alliez, and Mathieu Desbrun. Shapeshifter: 3d variations using multiscale and sparse point-voxel diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 605–617, 2025. 12

[47] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4460–4470, 2019. [48] John Willard Milnor. Morse theory. Number 51. Princeton university press, 1963. [49] Michael Moor, Max Horn, Bastian Rieck, and Karsten Borgwardt. Topological autoencoders. In International conference on machine learning, pages 7045–7054. PMLR, 2020. [50] Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. [51] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165–174, 2019. [52] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. [53] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652–660, 2017. [54] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748–8763. PmLR, 2021. [55] Xuanchi Ren, Jiahui Huang, Xiaohui Zeng, Ken Museth, Sanja Fidler, and Francis Williams. Xcube: Large-scale 3d generative modeling using sparse voxel hierarchies. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4209–4219, 2024. [56] Gernot Riegler, Ali Osman Ulusoy, and Andreas Geiger. Octnet: Learning deep 3d representations at high resolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3577–3586, 2017. [57] Haim Sawdayee, Amir Vaxman, and Amit H Bermano. Orex: Object reconstruction from planar cross-sections using neural fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20854–20862, 2023. [58] Tianchang Shen, Jacob Munkberg, Jon Hasselgren, Kangxue Yin, Zian Wang, Wenzheng Chen, Zan Gojcic, Sanja Fidler, Nicholas Sharp, and Jun Gao. Flexible isosurface extraction for gradient-based mesh optimization. ACM Transactions on Graphics (ToG), 42(4):1–16, 2023. [59] Tianqi Shen, Shaohua Liu, Jiaqi Feng, Ziye Ma, and Ning An. Topology-aware 3d gaussian splatting: Leveraging persistent homology for optimized structural integrity. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 6823–6832, 2025. [60] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110, 2023. [61] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multiview diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. [62] Suprosanna Shit, Johannes C Paetzold, Anjany Sekuboyina, Ivan Ezhov, Alexander Unger, Andrey Zhylka, Josien PW Pluim, Ulrich Bauer, and Bjoern H Menze. cldice-a novel topologypreserving loss function for tubular structure segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16560–16569, 2021. [63] Stefan Stojanov, Anh Thai, and James M Rehg. Using shape to categorize: Low-shot learning with an explicit shape bias. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1798–1808, 2021. 13

[64] Nico Stucki, Vincent Bürgin, Johannes C Paetzold, and Ulrich Bauer. Efficient betti matching enables topology-aware 3d segmentation via persistent homology. arXiv preprint arXiv:2407.04683, 2024. [65] Nico Stucki, Johannes C Paetzold, Suprosanna Shit, Bjoern Menze, and Ulrich Bauer. Topologically faithful image segmentation via induced matching of persistence barcodes. In International Conference on Machine Learning, pages 32698–32727. PMLR, 2023. [66] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. In European Conference on Computer Vision, pages 1–18. Springer, 2024. [67] Maxim Tatarchenko, Alexey Dosovitskiy, and Thomas Brox. Octree generating networks: Efficient convolutional architectures for high-resolution 3d outputs. In Proceedings of the IEEE international conference on computer vision, pages 2088–2096, 2017. [68] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion. In European Conference on Computer Vision, pages 439–457. Springer, 2024. [69] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, et al. Rodin: A generative model for sculpting 3d digital avatars using diffusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4563–4573, 2023. [70] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in neural information processing systems, 36:8406–8441, 2023. [71] Haohan Weng, Tianyu Yang, Jianan Wang, Yu Li, Tong Zhang, CL Chen, and Lei Zhang. Consistent123: Improve consistency for one image to 3d object synthesis. arXiv preprint arXiv:2310.08092, 2023. [72] Kailu Wu, Fangfu Liu, Zhihan Cai, Runjie Yan, Hanyang Wang, Yating Hu, Yueqi Duan, and Kaisheng Ma. Unique3d: High-quality and efficient 3d mesh generation from a single image. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. [73] Shuang Wu, Youtian Lin, Feihu Zhang, Yifei Zeng, Jingxi Xu, Philip Torr, Xun Cao, and Yao Yao. Direct3d: Scalable image-to-3d generation via 3d latent diffusion transformer. Advances in Neural Information Processing Systems, 37:121859–121881, 2024. [74] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaoou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1912–1920, 2015. [75] Jianfeng Xiang, Xiaoxue Chen, Sicheng Xu, Ruicheng Wang, Zelong Lv, Yu Deng, Hongyuan Zhu, Yue Dong, Hao Zhao, Nicholas Jing Yuan, et al. Native and compact structured latents for 3d generation. arXiv preprint arXiv:2512.14692, 2025. [76] Jianfeng Xiang, Zelong Lv, Sicheng Xu, Yu Deng, Ruicheng Wang, Bowen Zhang, Dong Chen, Xin Tong, and Jiaolong Yang. Structured 3d latents for scalable and versatile 3d generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21469–21480, 2025. [77] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191, 2024. [78] Chongjie Ye, Lingteng Qiu, Xiaodong Gu, Qi Zuo, Yushuang Wu, Zilong Dong, Liefeng Bo, Yuliang Xiu, and Xiaoguang Han. Stablenormal: Reducing diffusion variance for stable and sharp normal. ACM Transactions on Graphics (TOG), 43(6):1–18, 2024. 14

[79] Chongjie Ye, Yushuang Wu, Ziteng Lu, Jiahao Chang, Xiaoyang Guo, Jiaqing Zhou, Hao Zhao, and Xiaoguang Han. Hi3dgen: High-fidelity 3d geometry generation from images via normal bridging. arXiv preprint arXiv:2503.22236, 2025. [80] Tianjiao Yu, Xinzhuo Li, Yifan Shen, Yuanzhe Liu, and Ismini Lourentzou. Core3d: Collaborative reasoning as a foundation for 3d intelligence. arXiv preprint arXiv:2512.12768, 2025. [81] Tianjiao Yu, Xinzhuo Li, Yifan Shen, Onkar Susladkar, Yuanzhe Liu, Xiaona Zhou, and Ismini Lourentzou. Elsa3d: Elastic semantic anchoring for unified 3d understanding and generation. 2026. [82] Tianjiao Yu, Xinzhuo Li, Muntasir Wahed, Jerry Xiong, Yifan Shen, Ying Shen, and Ismini Lourentzou. Dreampartgen: Semantically grounded part-level 3d generation via collaborative latent denoising. 2026. [83] Tianjiao Yu, Vedant Shah, Muntasir Wahed, Ying Shen, Kiet A Nguyen, and Ismini Lourentzou. Part2 GS: Part-aware modeling of articulated objects using 3d gaussian splatting. arXiv preprint arXiv:2506.17212, 2025. [84] LAN Yushi, Shangchen Zhou, Zhaoyang Lyu, Fangzhou Hong, Shuai Yang, Bo Dai, Xingang Pan, and Chen Change Loy. Gaussiananything: Interactive point cloud flow matching for 3d generation. In The Thirteenth International Conference on Learning Representations, 2025. [85] Biao Zhang, Jiapeng Tang, Matthias Niessner, and Peter Wonka. 3dshape2vecset: A 3d shape representation for neural fields and generative diffusion models. ACM Transactions On Graphics (TOG), 42(4):1–16, 2023. [86] Chenyu Zhang, Yihao Luo, Yinzhe Wu, Choon Hwai Yap, and Guang Yang. Topologypreserving loss for accurate and anatomically consistent cardiac mesh reconstruction. arXiv preprint arXiv:2503.07874v1, 2025. [87] Longwen Zhang, Ziyu Wang, Qixuan Zhang, Qiwei Qiu, Anqi Pang, Haoran Jiang, Wei Yang, Lan Xu, and Jingyi Yu. Clay: A controllable large-scale generative model for creating high-quality 3d assets. ACM Transactions on Graphics (TOG), 43(4):1–20, 2024. [88] Ruowen Zhao, Zhengyi Wang, Yikai Wang, Zihan Zhou, and Jun Zhu. Flexidreamer: single image-to-3d generation with flexicubes. arXiv preprint arXiv:2404.00987, 2024. [89] Zibo Zhao, Zeqiang Lai, Qingxiang Lin, Yunfei Zhao, Haolin Liu, Shuhui Yang, Yifei Feng, Mingxin Yang, Sheng Zhang, Xianghui Yang, et al. Hunyuan3d 2.0: Scaling diffusion models for high resolution textured 3d assets generation. arXiv preprint arXiv:2501.12202, 2025. [90] Zibo Zhao, Wen Liu, Xin Chen, Xianfang Zeng, Rui Wang, Pei Cheng, Bin Fu, Tao Chen, Gang Yu, and Shenghua Gao. Michelangelo: Conditional 3d shape generation based on shapeimage-text aligned latent representation. Advances in neural information processing systems, 36:73969–73982, 2023. [91] Heming Zhu, Yu Cao, Hang Jin, Weikai Chen, Dong Du, Zhangye Wang, Shuguang Cui, and Xiaoguang Han. Deep fashion3d: A dataset and benchmark for 3d garment reconstruction from single images. In European Conference on Computer Vision, pages 512–530. Springer, 2020. [92] Afra Zomorodian and Gunnar Carlsson. Computing persistent homology. In Proceedings of the twentieth annual symposium on Computational geometry, pages 347–356, 2004.

15

A

Related Work

3D Generation. Early 3D generation methods commonly adapt pretrained 2D diffusion models to optimize each target asset through differentiable rendering or score distillation [52, 38, 70, 7, 43]. These methods reduce the need for large-scale 3D supervision, but they are often heavy in optimization and may inherit multi-view inconsistency from image priors. For image-conditioned generation, multi-view diffusion and reconstruction systems improve single-view consistency by predicting viewconsistent observations before 3D reconstruction [44, 45, 60, 88, 43, 71, 42, 72, 12, 68, 78, 40, 83, 23]. To improve scalability, later methods learn generative priors directly over compact 3D representations, including implicit fields, SDF latents, point or set latents, triplanes, and aligned image-text-shape latent spaces [79, 37, 31, 90, 69, 13, 33, 85, 6, 27, 82]. Recent systems further scale native 3D diffusion or rectified-flow transformers over learned latent tokens, Gaussian features, and sparse structured grids for high-quality conditioned asset generation [87, 66, 76, 89, 25]. In contrast, SILSA represents geometry as canonical slices along multiple axes, producing compact and spatially grounded latent tokens without explicitly predicting active voxels. Latent Representations for 3D Shapes. Designing compact but expressive latent representations is central to scalable 3D modeling [51, 47, 5, 85, 76, 8, 80]. Dense voxel grids provide explicit spatial structure but scale cubically with resolution [74, 14], motivating factorized representations such as triplanes and higher-dimensional plane decompositions, which encode 3D structure through axis-aligned feature planes rather than full volumetric grids [5, 20, 24]. Another line of work represents shapes with continuous implicit fields, including SDFs and occupancy functions, or compresses them into compact generative latents, volumetric codes, and unordered token sets [51, 47, 11, 13, 33, 85, 90, 87, 8]. These representations improve generative efficiency but often weaken explicit spatial correspondence between tokens and local geometry. A complementary direction preserves spatial locality through sparse or hierarchical voxels, reducing memory by modeling only occupied, adjacent, or progressively refined regions [56, 67, 55, 76, 25, 75]. OReX [57] shows that sparse planar cross-sections contain rich geometric information for neural-field reconstruction. Our representation differs by using axis-aligned slices not as external observations, but as a learned generative latent: a compact 3N -token layout that combines the spatial grounding of plane-based features with the bounded token count of set-based representations. Topology-Aware Learning. Topology-aware learning uses algebraic-topology tools, especially persistent homology [19, 92, 15, 22], to supervise structural properties that are poorly captured by pointwise or pixel-wise losses. In segmentation, persistent-homology and Betti-based objectives encourage predictions to match target connectivity and hole structure, while skeleton-based losses such as clDice provide efficient topology-preserving surrogates for curvilinear objects [30, 16, 62, 4, 65, 29]. Beyond output supervision, topology has also been used to regularize learned manifolds, as in Topological Autoencoders [49]. Recent work extends these ideas to 3D, including efficient Betti matching for volumetric segmentation, topology-constrained neural implicit reconstruction, and topology-aware reconstruction losses [64, 32, 86, 59, 28]. However, applying full volumetric persistent-homology supervision inside high-resolution generative training remains expensive, especially when topology must be evaluated repeatedly across decoded samples. SILSA makes topology supervision tractable by aligning the representation with axis-aligned cross-sections: persistent diagrams are matched within individual slices, while Betti transitions are matched across neighboring slices to preserve where topological events occur along each canonical axis.

B

Illustration of Multi-Axis Slice Topology

In Figure 5, we visualize how cross-sectional topology evolves as a 3D shape is sliced along the x-, y-, and z-axes. Each column corresponds to one slicing direction, with representative ground-truth cross-sections shown at the top and the induced topological events shown below. As the slicing plane moves through the object, each 2D cross-section induces connected components and holes. Blue intervals track connected components (β0 ), while red intervals track holes (β1 ) over slice depth. The endpoints of these intervals indicate topological transitions, such as a component appearing, a hole closing, or two regions merging. Our per-slice persistence loss matches the topology within each decoded cross-section, and our Betti-transition loss aligns the locations of these topological changes across adjacent slices. 16

X-axis Slice

Y-axis Slice

Z-axis Slice

Ground Truth

Ground Truth

Ground Truth

1

1

2

1 2

2

3 3

3

4

1

4

2

1

1

3

2

1

2

1

1

1

3

4

2

1 2

4

2

4 3

2

3

1

2

1 2

3

4

3

2

Figure 5: Topology signals for slice-wise supervision. For each canonical slicing direction, crosssections form a sequence over depth. Blue intervals denote connected components (β0 ) and red intervals denote holes (β1 ) that persist across ranges of slices. Our loss uses these signals in two ways: per-slice persistence matching supervises the topology within each cross-section, while Bettitransition matching supervises where components and holes appear, disappear, merge, or split across neighboring slices. Table 5: VAE reconstruction on open-surface shapes. Best and second best highlighted.

C

Model

CD ↓

F-Score@0.01 ↑

F-Score@0.005 ↑

IoU ↑

Betti-Err ↓

Trellis (SLAT) SparseFlex SILSA

0.07 0.05 0.04

99.71 100.00 100.00

91.18 93.07 93.21

96.84 98.42 98.71

0.01 0.01 0.00

Implementation Details

We train on Trellis-500K [76], curated from ObjaverseXL [18], ABO [17], 3DFUTURE [21], and HSSD [34]. Each mesh is normalized to a unit bounding box and encoded from Np =200K oriented surface samples. Unless otherwise stated, SILSA uses N =128 slice bins per axis, sliding-window width w=8, and latent dimension C=512, yielding 3N =384 slice tokens per shape. The decoder scatters latents into a 163 grid, applies sparse refinement and self-pruning upsampling to 2563 , and extracts meshes with differentiable Dual Marching Cubes [58, 35]. The Slice VAE is trained for 5 epochs with AdamW, learning rate 1×10−4 , batch size 16, and 8 NVIDIA A100 GPUs. We set λd =1.0, λn =0.5, λm =1.0, βKL =10−4 , and λtopo =0.1. For topology supervision, we use Ns =128 sampled cross-sections per axis and SDF-to-occupancy temperature κ=0.02. The rectified-flow transformer is conditioned on frozen DINOv2 image features [50]. It contains 24 layers, hidden dimension 1024, 16 attention heads, and a VAL resolution of D=16. We train for 300K steps using AdamW with learning rate 2×10−4 , cosine decay, batch size 256, and 8 NVIDIA A100 GPUs. At inference, we use 50 Euler steps and decode the predicted slice latents with the frozen VAE decoder.

D

Additional Results

Open Surface Evaluation. We further evaluate SliceVAE on the open-surface dataset DeepFashion3D [91]. As shown in Table 5, SILSA achieves the best reconstruction quality across geometric metrics, reducing CD from 0.05 to 0.04 over SparseFlex while matching its perfect F-Score@0.01 and improving F-Score@0.005 from 93.07 to 93.21. Topology-aware metrics saturate on this dataset because garments are topologically simple, so all methods that recover the rough surface achieve near-zero Betti-Err. The geometric improvements demonstrate that the slice-latent representation generalizes to open-surface shapes. Image-to-3D in the Wild. Figure 6 shows image-conditioned generations on diverse in-the-wild examples. SILSA recovers plausible 3D structure from a single view and maintains consistency across rendered viewpoints. The results are strongest on objects whose geometry is difficult for 17

GT

Figure 6: Additional image-to-3D results. compact global latents, including chairs with legs and armrests, drones with thin propeller supports, motorcycles with wheels and handles, and flowers with layered petals. These examples show that the proposed cross-axis slice representation provides enough local structure to reconstruct fine details while still producing globally coherent 3D assets.

E

Ablations

Slice representation. Table 6 studies how the slice resolution and sliding-window width affect VAE reconstruction. Increasing the number of slices improves reconstruction quality from N =64 to N =256, since finer slice bins expose more local geometry and reduce topological ambiguity. However, the gains saturate beyond N =128: N =256 and N =512 slightly improve CD and IoU, but require 2× and 4× more tokens, while N =1024 further increases token count and worsens Betti-Err, suggesting that overly fine slicing fragments cross-sectional evidence. We therefore use N =128 as the default because it provides the best efficiency-fidelity trade-off with only 384 tokens. The 18

Table 6: Ablation on slice representation. We ablate the number of slices per axis N and the sliding-window width w. Default settings are N =128, w=8, and three canonical axes, yielding 3N =384 slice tokens. Reported on the VAE reconstruction task.

#Tokens

CD ↓

F@0.01 ↑

IoU ↑

Betti-Err ↓

192 288 384 768 1536 3072

0.82 0.67 0.59 0.57 0.55 0.56

94.36 95.81 96.79 97.02 97.24 97.18

88.24 91.27 93.01 93.28 93.51 93.42

2.86 2.04 1.58 1.53 1.63 1.77

Window width w at N =128 w=1 (no overlap) 384 w=2 384 w=4 384 w=8 (default) 384 w=16 384

0.73 0.68 0.63 0.59 0.62

94.18 95.09 96.14 96.79 96.26

86.95 89.32 91.85 93.01 92.04

3.24 2.46 1.91 1.58 1.83

Window width w at N =256 w=1 (no overlap) 768 w=2 768 w=4 768 w=8 768 w=16 768

0.64 0.61 0.58 0.57 0.59

95.67 96.34 96.86 97.02 96.74

90.18 91.47 92.92 93.28 92.80

2.35 2.03 1.67 1.53 1.61

Variant Number of slices N N =64 N =96 N =128 (default) N =256 N =512 N =1024

window-width ablation shows that overlap is critical: w=1 and w=2 lack sufficient context and produce higher Betti error, while w=8 gives the strongest overall balance. Increasing the window to w=16 slightly degrades performance, likely because excessive aggregation smooths local structures and weakens slice-level specificity. Topology Loss. Table 7 studies the contribution of the slice-wise topology-preserving objective. Removing Ltopo weakens structural preservation because rendering losses alone do not explicitly penalize broken components, filled holes, or incorrect connectivity changes across slices. Using only LPH improves the topology of individual cross-sections, but does not directly constrain where topological events occur along the slicing direction. Conversely, using only Ltrans encourages event locations to align across depth, but provides weaker supervision for the detailed topology within each slice. The full objective combines these complementary signals and gives the best balance between surface fidelity and topological correctness. The ablation over λtopo further shows that a moderate topology weight is preferable: too small a weight provides limited structural supervision, while too large a weight can over-constrain the decoder and reduce geometric fidelity. Increasing the number of supervised cross-sections improves the coverage of topological events, with the default Ns =128 providing the strongest supervision without changing the compact latent layout. Number of Axes. Table 8 ablates the number of canonical slicing axes used by the SliceVAE while approximately matching the total token budget. Single-axis variants degrade substantially regardless of slicing direction: even when given N =384 slices to match the default token count, the model lacks cross-sectional evidence orthogonal to its slicing direction, leading to severe topological errors (Betti-Err of 5.92 for z-only and 6.18 for x-only). Reducing the budget to N =128 amplifies the gap further, confirming that single-axis representations cannot recover what is missing in their orthogonal directions. Two-axis variants close most of the gap, with (x, z) slightly outperforming (x, y) since orthogonal vertical and horizontal slicing captures more complementary structure for typical upright objects. However, both two-axis configurations still trail the full three-axis design, particularly on Betti-Err (2.61–2.74 vs. 1.58), indicating that the third axis specifically reinforces topological consistency by exposing structures that any two cross-sectional views jointly underdetermine. 19

Table 7: Ablation of the slice-wise topology-preserving loss. All variants are evaluated on VAE reconstruction. Default settings are shaded. Variant

CD ↓

F@0.01 ↑

F@0.005 ↑

IoU ↑

Betti-Err ↓

Loss components No Ltopo Only LPH Only Ltrans Full Ltopo (default)

0.78 0.72 0.69 0.59

93.84 94.51 95.03 96.79

78.16 79.72 80.46 84.03

81.76 84.41 87.16 93.01

4.43 2.91 2.87 1.58

Topology loss weight λtopo λtopo =0.01 0.66 λtopo =0.1 (default) 0.59 λtopo =0.5 0.62 λtopo =1.0 0.67

95.62 96.79 96.31 95.41

82.14 84.03 83.27 81.92

90.37 93.01 92.42 90.86

2.24 1.58 1.73 1.96

Number of cross-sections Ns per axis Ns =32 0.68 95.07 Ns =64 0.63 96.02 Ns =128 (default) 0.59 96.79

81.03 82.75 84.03

88.94 91.48 93.01

2.36 1.87 1.58

Table 8: Ablation on the number of canonical axes. Default setting is shaded. Axes

N

#Tokens

CD ↓

IoU ↑

Betti-Err ↓

Single axis 1 axis (z only) 1 axis (z only) 1 axis (x only) 1 axis (x only)

z z x x

384 128 384 128

384 128 384 128

1.34 1.87 1.41 1.94

76.43 71.29 75.18 70.42

5.92 7.83 6.18 8.07

Two axes 2 axes (x, y) 2 axes (x, z)

x, y x, z

192 192

384 384

0.81 0.79

88.46 88.91

2.74 2.61

Three axes 3 axes (default)

x, y, z

128

384

0.59

93.01

1.58

Variant

VAL Update Mechanism. Table 9 ablates how slice tokens write to the Volumetric Anchor Lattice. Replacing gated writes with full overwrite causes each transformer block to clobber the accumulated cross-axis evidence with the latest token’s contribution, breaking the multi-block coordination that makes VAL effective (CD 0.66, Betti-Err 1.91). Switching to additive writes preserves prior evidence but lets magnitudes accumulate without channel-wise selectivity, which improves over overwrite but still trails the gated variant (CD 0.63, Betti-Err 1.74). The default gated write achieves the best results across all metrics by allowing the model to learn, per channel, how much existing VAL content to retain versus replace as new slice evidence arrives. This selectivity is what enables the VAL to function as a stable shared workspace across transformer blocks rather than as a noisy buffer.

F

Discussion

3D Topology Preservation. Topology refers to the structural properties of a shape that remain invariant under continuous deformation: the number of connected components, the presence and count of holes, and the way these structures relate across the object. For 3D shapes, these properties are formalized through Betti numbers, where β0 counts connected components, β1 counts loops or tunnels, and β2 counts enclosed voids. Unlike point-wise geometric metrics such as Chamfer distance or surface error, topological correctness captures whether a reconstructed shape preserves the qualitative structure of the original, whether a chair has four separable legs rather than three fused ones, whether a wheel retains its central opening rather than filling in, and whether a railing’s spokes remain individually disconnected from the surrounding frame. These distinctions matter because shapes with low surface error can still be structurally wrong: a generated mesh that fills a hole, breaks 20

Table 9: Ablation on the VAL update mechanism. Default setting is shaded. Variant

CD ↓

F@0.01 ↑

IoU ↑

Betti-Err ↓

Overwrite (no gating) Additive write (no gating) Gated write (default)

0.66 0.63 0.59

95.61 96.12 96.79

90.86 92.04 93.01

1.91 1.74 1.58

a thin support, or merges two nearby parts will register only small per-vertex deviations from the ground truth while fundamentally misrepresenting what the object is. Persistent homology [19, 92] provides a principled way to quantify and supervise these structural properties during learning. By tracking how topological features appear and disappear across a filtration of the shape — for example, sweeping a level set through an SDF — persistent homology produces a multi-scale signature that records each feature’s birth, death, and persistence. Topological features that persist across a wide range of filtration values correspond to robust structures, while short-lived features correspond to noise. Loss functions built on persistent homology and Bettinumber matching have proven effective for 2D segmentation tasks involving thin or branching structures [30, 16, 62, 65], and recent work has extended these ideas to 3D segmentation and implicit reconstruction [32, 86]. However, applying full volumetric topology supervision to high-resolution 3D generation remains computationally prohibitive, since persistence diagrams must be recomputed across many decoded samples and at fine spatial resolution. SILSA addresses this by exploiting the slice-based structure of its latent representation: persistence is matched within individual crosssections, and Betti transitions are aligned across neighboring slices, supervising topology at tractable per-slice cost while still capturing how connectivity evolves through the shape. Limitations. Like most learning-based 3D generation methods, SILSA’s performance depends on the diversity and scale of the training distribution, and objects with structural patterns far outside this distribution may be reconstructed with reduced fidelity. Generation quality is also influenced by the quality of the input image, with ambiguous or low-information views potentially yielding less faithful 3D structure. Scaling to broader data sources is a promising direction for future work. Broader Impact. Our method contributes to high-resolution image-to-3D generation, with positive applications in content creation, design, education, and simulation, where it lowers the barrier to producing 3D assets. The improved efficiency of our slice-based representation also makes high-resolution 3D generation more accessible to researchers with limited computational resources. As with other generative models, advances in 3D generation carry risks including unauthorized 3D replicas of real objects and the displacement of manual modeling tasks, which downstream applications should address through appropriate safeguards.

21

Related documents

• SILSA: Sliding-Window Slice Latents for Topology-Preserving High-Resolution 3D Generation#1157549
• Stereological reconstructions of 3D cellular microstructures by combining adversarial learning and Voronoi tessellations.#185880
• Zero-Shot Novel Depth Synthesis Using 3D Foundation Models Scene Representations#927039
• FLOWR.ROOT – A flow matching-based foundation model for joint multi-purpose structure-aware 3D ligand generation and affinity prediction#485073
• Rethinking the State Update Gate for Long-Sequence Recurrent 3D Reconstruction#1131538