💻 computer science

LaS-Comp: Zero-shot 3D Completion with Latent-Spatial Consistency

LaS-Comp is a zero-shot, training-free framework that leverages 3D foundation model priors through a two-stage design of explicit replacement and implicit refinement to achieve high-quality, category-agnostic 3D shape completion, validated by the new Omni-Comp benchmark.

Original authors: Weilong Yan, Haipeng Li, Hao Xu, Nianjin Ye, Yihao Ai, Shuaicheng Liu, Jingyu Hu

Published 2026-03-19

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Weilong Yan, Haipeng Li, Hao Xu, Nianjin Ye, Yihao Ai, Shuaicheng Liu, Jingyu Hu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are an artist trying to finish a sculpture, but you only have a few broken shards of clay in your hand. Your goal is to recreate the entire statue, making sure the new parts fit perfectly with the old shards and look like a real, complete object.

This is the challenge of 3D Shape Completion. For a long time, computers struggled with this. If you showed them a broken chair, they might guess it's a table, or they might smooth over the broken parts so much that the chair looks like a blob. They often needed to be trained on thousands of "broken chair + whole chair" pairs, which limits them to only what they've seen before.

Enter LaS-Comp, a new "super-artist" AI that can finish any 3D object it sees, even if it has never seen that specific object before. Here is how it works, explained simply:

The Problem: The "Translation" Glitch

The paper starts with a clever observation. Modern AI models that generate 3D shapes work in a "secret code" (called Latent Space). Think of this like a language where the AI speaks.

The Issue: If you take a broken chair and ask the AI to translate it into its secret code, and then take a whole chair and translate that, the "secret code" for the matching parts (the legs, the seat) doesn't actually match! It's like if you wrote "cat" in English and "cat" in French, but the AI thought they were completely different words.
The Result: If the AI tries to just "fill in the blanks" using this secret code, it gets confused and creates a mess because the code for the broken part doesn't line up with the code for the whole part.

The Solution: LaS-Comp's Two-Step Dance

The authors created a framework called LaS-Comp (Latent–Spatial Consistency). Instead of trying to fix the secret code directly, they use a two-step dance to keep the original shape safe while letting the AI do its magic.

Step 1: The "Sticky Tape" (Explicit Replacement Stage)

Imagine you have a broken vase. Before asking the AI to guess the rest, you take the actual broken shards and physically tape them back onto the new vase you are about to build.

What the AI does: It takes the "secret code" of the new shape and forcibly swaps out the parts that correspond to your broken shards with the actual data from your shards.
Why it helps: This guarantees that the AI cannot accidentally change your original input. The broken chair leg stays exactly where it is. It's like saying, "I know this part is real; don't touch it."

Step 2: The "Smoothie Blender" (Implicit Alignment Stage)

Now, you have a vase where the taped-on shards look a bit jagged against the new clay. There's a rough seam.

What the AI does: It performs a quick, tiny "polish." It looks at the edge where the real shard meets the new clay and gently nudges the math to make the transition smooth. It's like using a fine sandpaper to blend the new clay into the old shard so you can't tell where one ends and the other begins.
Why it helps: This prevents the "glitchy" look where the new parts look like they are floating or disconnected from the old parts.

Why This is a Big Deal

Zero-Shot (No Training Required): You don't need to teach this AI with thousands of examples. It uses a "pre-trained" brain (a 3D foundation model) that already knows what chairs, cars, and animals generally look like. It just applies that knowledge to your specific broken piece.
Works on Anything: Whether you have a single photo of a chair, a random chunk of a robot, or a missing piece of a statue, this method handles it. Previous methods often failed if the broken piece was too weird or if the object was partially hidden.
Speed: It's incredibly fast, finishing a shape in about 20 seconds.

The New "Test Drive" (Omni-Comp)

The authors realized that old tests were too easy. They mostly used simple "single photo" scans. To prove their method is truly tough, they built a new test called Omni-Comp.

Imagine a driving test. Old tests were like driving on a straight, empty road.
Omni-Comp is like driving in a storm, with potholes, random obstacles, and missing road signs. It tests the AI on:
- Random Crops: Like someone took a bite out of the object.
- Semantic Parts: Like showing only the wheels of a car, with the rest missing.
- Real-World Noise: Messy, real-life scans from robots and cameras.

The Verdict

In simple terms, LaS-Comp is like a master restorer who doesn't just guess what the missing piece looks like. Instead, they:

Lock the original broken pieces in place so they don't move.
Grow the rest of the object around them using their deep knowledge of how the world works.
Polish the seams until the whole thing looks like it was made in one piece.

The result? A computer that can look at a broken toy, a half-scanned car, or a missing statue part, and instantly "hallucinate" the rest of it with high accuracy, without needing to be taught how to do it first.

1. Problem Statement

3D Shape Completion aims to reconstruct a complete 3D object from partial observations (e.g., single-view scans, occluded regions, or missing semantic parts). While critical for robotics, autonomous driving, and AR/VR, existing methods face significant limitations:

Supervised Methods: Rely on paired partial-complete datasets, limiting generalization to unseen object categories.
Unsupervised/Generative Prior Methods: Recent zero-shot approaches leverage large pre-trained models but often assume the partial input can be rendered into at least one "complete" 2D view. When the missing geometry is visible from all viewpoints (e.g., a hollowed-out object or severe occlusion), these rendering-based assumptions fail, leading to degraded results.
Latent-Spatial Gap: A fundamental challenge exists when using latent-generative 3D foundation models (which map shapes to a compact latent space). Even if a partial input and a ground-truth shape share identical geometry in overlapping regions, their latent encodings differ significantly. Directly using partial latent codes as conditions for generation leads to unreliable completions due to this domain gap.

2. Methodology: LaS-Comp

The authors propose LaS-Comp, a zero-shot, category-agnostic framework that bridges the gap between the latent and spatial domains of pre-trained 3D foundation models. It operates without training on specific datasets and is compatible with various latent-based generative models (e.g., TRELLIS, Direct3D-S2).

The framework utilizes a complementary two-stage design within an iterative denoising process:

A. Explicit Replacement Stage (ERS)

Goal: Ensure strict fidelity to the observed partial geometry ( $S_p$ ).
Mechanism:
1. Dual-Branch Decomposition: The generation process is split into a "clean branch" (estimating noise-free latents) and a "noisy branch" (handling stochasticity).
2. Spatial Injection: The partial input $S_p$ is voxelized and used to create a binary mask $M$ . The generative model's prediction is explicitly replaced with the observed geometry $S_p$ in the spatial domain where $M=1$ .
3. Partial-aware Noise Schedule (PNS): To handle the asymmetry between known and unknown regions, the noise schedule is modulated. Observed regions receive minimal stochastic disturbance to preserve geometry, while missing regions receive pure Gaussian noise to encourage diverse exploration of plausible completions.
4. Latent Reconstruction: The spatially corrected shape is re-encoded into the latent space ( $x^*_t$ ), ensuring the latent representation explicitly contains the observed geometry.

B. Implicit Alignment Stage (IAS)

Goal: Resolve boundary inconsistencies and ensure smooth transitions between observed and synthesized regions.
Mechanism:
1. Geometry-Alignment Loss: A loss function ( $L_{align}$ ) is computed between the decoded prediction and the partial input only in the observed regions (using Binary Cross-Entropy).
2. Gradient-Based Refinement: Instead of updating model weights, the latent feature is optimized via a single-step gradient descent to minimize this loss.
3. Result: This produces a refined latent ( $x^{aligned}_{0|t}$ ) that aligns the generative prior with the observed data, smoothing artifacts before the next denoising step.

3. Key Contributions

LaS-Comp Framework: A novel zero-shot approach that solves the latent-spatial inconsistency problem by combining Explicit Replacement (for fidelity) and Implicit Alignment (for coherence).
Training-Free & Efficient: The method requires no fine-tuning of the foundation models. It is highly efficient, completing a shape in approximately 20 seconds (over 3× faster than existing zero-shot methods).
Omni-Comp Benchmark: The authors introduce a comprehensive new benchmark containing 30 diverse objects (real-world scans from Redwood/YCB and synthetic shapes) with three distinct partial patterns:
- Single Scan (standard depth/LiDAR).
- Random Crop (arbitrary occlusions).
- Semantic Part (missing specific semantic components).
- Significance: This addresses the limitation of previous benchmarks that only tested single-view scans.

4. Experimental Results

The method was evaluated on Redwood, Synthetic, KITTI, ScanNet, and the new Omni-Comp benchmark.

Quantitative Performance:
- Redwood/Synthetic: Outperforms state-of-the-art zero-shot methods (ComPC, GenPC, SDS-Comp) by significant margins (e.g., 27.2% improvement in Chamfer Distance over ComPC).
- Omni-Comp: Demonstrates robustness across all partial patterns. It achieves a 49.6% improvement in CD and 39.4% in EMD over the best baseline (ComPC) on average.
- Fidelity: Achieves superior Unidirectional Chamfer Distance (UCD) and Hausdorff Distance (UHD), proving it preserves input geometry better than competitors.
Qualitative Performance:
- Successfully handles challenging cases where baselines fail, such as objects with missing semantic parts or severe occlusions where no "complete" 2D view exists.
- Generates structurally plausible and geometrically detailed completions without distorting visible regions.
Diversity: The method supports both unconditional and text-guided completion, generating diverse hypotheses (high TMD scores) while maintaining geometric consistency.

5. Significance and Impact

Overcoming the Rendering Assumption: LaS-Comp is the first framework to fully leverage latent-generative 3D foundation models without relying on the assumption that a partial input can be rendered into a complete 2D image. This makes it applicable to highly irregular and occluded real-world scenarios.
Bridging the Latent-Spatial Gap: By explicitly injecting spatial geometry into the latent generation process, the paper provides a robust solution to the domain gap that has hindered zero-shot 3D completion.
Practical Deployment: The training-free nature and compatibility with existing foundation models make it immediately deployable for real-world applications like robotic grasping and autonomous navigation, where object categories are open-ended and data is scarce.
New Evaluation Standard: The introduction of Omni-Comp sets a new standard for evaluating 3D completion, moving beyond simple single-view scans to test true generalization across diverse occlusion patterns.