Pixal3D is an open-source single-image 3D generation model from Tencent ARC Lab, developed with Tsinghua University and Victoria University of Wellington, and accepted to SIGGRAPH 2026. It turns one photograph into a high-fidelity textured 3D asset exported as a GLB mesh. The project's central claim is a change in how image information reaches the 3D representation: instead of loosely injecting image features through attention, as many prior methods do, Pixal3D explicitly lifts pixel features into 3D by back-projection, establishing direct pixel-to-3D correspondences that the authors argue push output quality toward reconstruction-level fidelity. ## Pixel-Aligned Generation The idea behind the name is straightforward but consequential. In a typical image-to-3D pipeline, a network attends over image tokens and hopes the relevant detail survives into the geometry; fine structures and surface detail often blur in the process. Pixal3D instead back-projects pixel features along known camera rays so that each part of the generated shape is anchored to specific pixels in the input. This explicit correspondence is what lets the model recover detailed geometry and physically based rendering (PBR) textures rather than a smoothed approximation, narrowing the gap between generation from a single view and true multi-view reconstruction. ## Two Backbones, Two Branches The repository is organized around two implementations. The `main` branch is the latest version, an improved implementation built on the TRELLIS.2 backbone that the authors report gives better performance. The `paper` branch preserves the original implementation on the Direct3D-S2 backbone and corresponds exactly to the numbers reported in the SIGGRAPH 2026 paper, so researchers who need to reproduce published results can check out that branch specifically. This separation is a practical touch: it lets the project keep improving the shipping model while keeping the paper's claims verifiable against frozen code. ## Running It in Practice Getting a mesh out of Pixal3D is a single command — `python inference.py --image input.png --output output.glb` — once the environment is set up on top of the TRELLIS.2 base install, with additional dependencies including the `natten` neighborhood-attention kernels and a `utils3d` wheel. The default pipeline runs at 1536 resolution, and a low-VRAM mode loads models on demand to cut peak memory, dropping to 1024 by default but overridable. For users who do not want to install anything, the team hosts an interactive Gradio demo on Hugging Face Spaces, and the same Gradio app can be run locally with `python app.py`, including the low-VRAM path. An `ATTN_BACKEND=sdpa` option lets people without FlashAttention still run the model. ## An Open Release Pixal3D arrived as a fairly complete open release rather than a teaser. Over May 2026 the authors published inference code and the online demo, the improved TRELLIS.2-based version, and a training code and data-preparation toolkit, all under the permissive MIT license. Releasing the training pipeline alongside inference is notable for a 3D generation paper: it lets other groups retrain or fine-tune on their own data instead of being limited to the released checkpoint, which matters in a field where high-quality 3D training data is scarce and expensive to assemble. ## Trade-offs and Limitations The constraints are those of research-grade 3D generation. The pipeline is built on the TRELLIS.2 environment and CUDA-specific kernels like `natten`, so setup is more involved than a pip install and effectively requires an NVIDIA GPU; the low-VRAM mode helps but does not remove the accelerator requirement. Full 1536-resolution generation is memory-hungry, and results depend on the quality and clarity of the single input image — occluded or ambiguous regions must be hallucinated because only one view is available. The two-branch layout, while good for reproducibility, means newcomers must be deliberate about which backbone they are running. ## Who Should Use This Pixal3D is aimed at researchers in 3D generation and at technical artists and developers building asset pipelines who need textured, game-ready GLB meshes from a single reference image. The pixel-aligned approach makes it especially interesting where fidelity to the input matters — product visualization, prototyping, and any workflow where the generated model should look like the specific object photographed rather than a generic lookalike. Teams without GPU infrastructure will find the Hugging Face demo the fastest way to evaluate whether the quality fits their needs before committing to a local install.