TencentARC's Pixal3D Turns a Single Photo Into a 4K 3D Asset

Pixal3D, TencentARC's SIGGRAPH 2026 image-to-3D model, is now officially repackaged for ComfyUI by the Comfy-Org team. The model has already racked up over 5,000 downloads on Hugging Face, and the community has wasted no time building custom node packs around it. If you've been frustrated by the blurry, clay-like outputs from previous image-to-3D tools, this one is worth your attention.

The Problem Every 3D Generator Has Had

Whenever a model attempts to infer 3D structure from a 2D image, it has to guess the depth, the occluded angles, and the structural relationships between pixels. Most architectures relied on feeding compressed image tokens to an attention mechanism and hoping the network learned spatial geometry along the way. The result was usually a 3D asset that looked great from a distance but fell apart upon closer inspection. Fine details like text on a label, stitching on a shoe, or mechanical parts would get washed out or hallucinated entirely.

Models often suffer from feature bleeding, where the texture of the background or neighboring objects bleeds onto the target asset. The root cause is architectural: image encoders like CLIP or DINOv2 flatten rich 2D images into one-dimensional token sequences, destroying the native spatial structure that geometry depends on.

Pixel Alignment: Math Instead of Guessing

Pixal3D generates high-fidelity 3D assets from a single image. Unlike previous methods that loosely inject image features via attention, Pixal3D explicitly lifts pixel features into 3D through back-projection, establishing direct pixel-to-3D correspondences. This enables near-reconstruction-level fidelity with detailed geometry and PBR textures.

Back-projection (also called explicit lifting) is the key innovation here. The idea is borrowed from classical computer vision: if you know the camera angle from which a photo was taken, you can draw a mathematical ray from the camera lens, through each pixel, and out into 3D space. By establishing a direct, mathematical correspondence between the 2D pixel and the 3D coordinate, the model achieves near-reconstruction-level fidelity. In PyTorch terms, this boils down to a single well-optimized call: