TencentARC/Pixal3D

[SIGGRAPH 2026] Pixal3D: Pixel-Aligned 3D Generation from Images

What it solves

Pixal3D addresses the lack of precision in 3D asset generation from single images. While previous methods often loosely inject image features, Pixal3D ensures high-fidelity geometry and PBR textures by establishing direct pixel-to-3D correspondences.

How it works

The system uses a three-stage cascade to progressively increase resolution:

  1. Sparse Structure: Generates initial coarse structure (32 to 64 resolution).
  2. Shape: Refines the geometry (256 to 1024 resolution).
  3. Texture: Adds detailed PBR textures (256 to 1024 resolution).

It achieves this by explicitly lifting pixel features into 3D through back-projection, rather than relying on attention mechanisms. It can operate on a single image or multiple views of the same object using a separate set of multi-view weights.

Who it’s for

  • 3D Artists and Designers: To quickly generate high-quality 3D assets from 2D images.
  • AI Researchers: Those working on 3D generation and pixel-aligned projection conditioning.
  • Developers: Users looking to integrate 3D generation into their workflows via the provided Gradio web demo or ComfyUI extensions.

Highlights

  • Pixel-Aligned Projection: Establishes direct correspondences between pixels and 3D space for near-reconstruction-level fidelity.
  • Multi-View Support: Ability to condition the generation process on multiple images of an object for better accuracy.
  • PBR Textures: Generates assets with detailed Physically Based Rendering textures.
  • Cascaded Training: A structured three-stage pipeline that scales from sparse structure to high-resolution texture.
  • Low-VRAM Mode: Includes an option to reduce peak memory usage for accessibility on consumer hardware.

Related

  • Project
    IBM/AssetOpsBench

    AssetOpsBench is an open‑source benchmark/framework for building, orchestrating, and evaluating LLM‑based AI agents that operate on industrial asset‑management data (sensors, work orders, failure modes, etc.). It provides domain‑specific MCP tool servers, several ReAct‑style agent back‑ends, 141+ realistic scenarios, and a multi‑dimensional evaluation pipeline used in KDD/AAAI/NeurIPS competitions.

  • Project
    PrismML-Eng/Bonsai-demo

    A demo repository for running Bonsai and Ternary-Bonsai language models locally, featuring extreme 1-bit and 2-bit quantization for high-performance inference on consumer hardware.

  • Project
    nv-tlabs/lyra

    A series of open generative 3D world models from NVIDIA that create 3D and 4D scenes from a single image or video.

  • Project
    MoonshotAI/Kimi-K2.5

    Kimi K2.5 is an open-source, native multimodal agentic model with a 1T parameter MoE architecture that integrates vision and language understanding for complex task execution.