# InpaintSLat: Inpainting Structured 3D Latents via Initial Noise Optimization

Source: [InpaintSLat: Inpainting Structured 3D Latents via Initial Noise Optimization](<https://www.sota2.com/research/paper/69f827e7c0f56ed8430888ee>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2026-05-05T05:00:04.460Z

Authors: Jaeyoung Chung, Suyoung Lee, Kyoung Mu Lee

Published: 2026-05-01T13:45:25.000Z

## Abstract

We present a training-free approach for controllable 3D inpainting based on initial noise optimization. In the structured 3D latent diffusion framework, we observe that the underlying geometric structure is established during the early stages of the diffusion process and exhibits high sensitivity to the initial noise. Such characteristics compromise stability in tasks like inpainting and editing, where the model must ensure strict alignment with the existing context while synthesizing a new structure. In this paper, we introduce a strategy to optimize the initial noise within the structured 3D latent diffusion framework, ensuring high-fidelity 3D inpainting. Specifically, we update the initial noise by leveraging a backpropagation approximation grounded in the rectified flow model, with the spectral parameterization specially designed for robust and efficient structured 3D latent optimization. Experiments demonstrate consistent improvements in contextual consistency and prompt alignment over representative training-free inpainting baselines, establishing initial noise control as an independent dimension for 3D inpainting, orthogonal to conventional sampling trajectory manipulation.

## Recorded benchmark results

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

- [3D Reconstruction on Toys4k (Preserved Part)](<https://www.sota2.com/research/sota/3d-reconstruction-on-toys4k-preserved-part.md>): 3D Reconstruction; Toys4k (Preserved Part); Appearance PSNR: 25.46.

- [Appearance Reconstruction on ABO](<https://www.sota2.com/research/sota/appearance-reconstruction-on-abo.md>): Appearance reconstruction; ABO; PSNR: 25.14.

- [Appearance Reconstruction on Toys4k Preserved Part](<https://www.sota2.com/research/sota/appearance-reconstruction-on-toys4k-preserved-part.md>): Appearance reconstruction; Toys4k (Preserved Part); PSNR: 22.8.

- [Inpainting on ABO (test)](<https://www.sota2.com/research/sota/inpainting-on-abo-test.md>): Inpainting; ABO (test); CLIP Score: 28.36.

- [Normal Map Geometry Reconstruction on ABO](<https://www.sota2.com/research/sota/normal-map-geometry-reconstruction-on-abo.md>): Normal Map Geometry Reconstruction; ABO; PSNR: 27.3252.

- [Point Cloud Geometry Reconstruction on ABO](<https://www.sota2.com/research/sota/point-cloud-geometry-reconstruction-on-abo.md>): Point Cloud Geometry Reconstruction; ABO; Chamfer Distance (L1): 0.541.

- [3D Inpainting on Toys4k Inpainting Part](<https://www.sota2.com/research/sota/3d-inpainting-on-toys4k-inpainting-part.md>): 3D Inpainting; Toys4k Inpainting Part; CLIP Score: 30.42.

## Original sources and implementation

[Original paper](<https://arxiv.org/abs/2605.00664>)

[PDF](<https://arxiv.org/pdf/2605.00664.pdf>)
