RunningHub Pixal3D Image to 3D

This is the node that actually does the thing. Feed it one image and it runs TencentARC's Pixal3D, an image-to-3D pipeline that reconstructs a textured, PBR-shaded 3D asset from a single photo or generated image. Everything else in the pack is plumbing around it: the loader feeds it a model bundle, and its asset output is what Save GLB turns into a file on disk.

What you're really getting

Set expectations before you spend an evening on the install. Image-to-3D is the one corner of 3D generation that genuinely works, and Pixal3D is among the best open entries in it - the community read it on release as a clear step up from raw TRELLIS.2. But "high-fidelity 3D asset" means aesthetic quality, not game-ready topology. You'll get a mesh with machine-generated triangle soup and UVs; it's brilliant for static props, renders, and 3D printing, and it is not rig-ready. If your destination is a game engine with animation, budget for a retopology pass.

Pixal3D also ships under a research license - academic use only, and explicitly "not intended for use within the European Union." Check the license before building anything commercial on it.

How the pipeline works

Pixal3D runs three flow-matching stages in sequence, and this node exposes every one of their knobs:

1. Sparse structure (ss_*) - sketches the coarse voxel scaffold of the object.
2. Shape SLAT (shape_slat_*) - generates the structured latent that becomes the actual mesh.
3. Texture SLAT (tex_slat_*) - paints the surface with PBR material channels.

That SLAT design is straight from the TRELLIS lineage - Pixal3D is built on that stack, right down to exporting through TRELLIS.2's o_voxel code. Before sampling, a MoGe model estimates the camera from your image, and the image is conditioned through DINOv3 features plus BiRefNet background removal and NAF upsampling - which is why the loader drags in five separate model folders. None of that is your problem; it all happens inside the node.

The inputs that actually matter

The required list is intimidating - 14 knobs - but you'll touch maybe three of them:

• resolution: 1024 or 1536. The generation grid size. 1024 is the sane default; 1536 buys detail and eats VRAM and time.
• ss_sampling_steps / shape_slat_sampling_steps / tex_slat_sampling_steps: the biggest quality-versus-speed levers. 12 is the shipped default. Drop all three to 1 for a smoke test - you'll know in a minute whether the pipeline works instead of waiting twenty.
• seed: reproducibility, and if a generation lands weird geometry, bumping it is the first thing to try.

Everything else - the per-stage guidance strengths, rescales, and rescale_t values - are CFG-style knobs; leave them at the tuned defaults. Two optional inputs are worth knowing: mask, which overrides the automatic background removal (wire in a MASK if the BiRefNet cutout is wrong), and max_num_tokens, the sparse-token cap for high-res stages - lower it if 1536 OOMs.

Outputs

• asset (PIXAL3D_ASSET) - wire this into RunningHub Pixal3D Save GLB; it carries the mesh and everything the exporter needs.
• camera_info (STRING) - the estimated camera parameters (angle, distance, mesh scale) as text, for downstream preview nodes.

Installing the whole stack

This node inherits the entire pack's weight. From the README:

cd ComfyUI/custom_nodes
git clone https://github.com/RH-RunningHub/ComfyUI_RH_Pixal3D.git
cd ComfyUI_RH_Pixal3D
pip install -r requirements.txt

The requirements file only pulls MoGe and huggingface_hub - the scary stuff is manual. NATTEN needs a matching wheel plus libnatten:

pip uninstall -y natten
pip install natten==0.21.6+torch2100cu128 -f https://whl.natten.org
python -c "import natten; print(natten.__version__, natten.HAS_LIBNATTEN)"   # True

Then the CUDA extensions (flash_attn, flex_gemm, o_voxel, nvdiffrast, nvdiffrec_render) need wheels or source builds matched to your exact Python/PyTorch/CUDA. The README's install notes are the authority here - this is the corner where "it won't import" eats weekends.

The nodes never download models during execution. From your ComfyUI directory:

mkdir -p models/Pixal3D models/MoGe/moge-2-vitl models/DINOv3/dinov3-vitl16-pretrain-lvd1689m models/BiRefNet/RMBG-2.0 models/NAF
hf download TencentARC/Pixal3D --local-dir models/Pixal3D
hf download Ruicheng/moge-2-vitl --local-dir models/MoGe/moge-2-vitl
hf download facebook/dinov3-vitl16-pretrain-lvd1689m --local-dir models/DINOv3/dinov3-vitl16-pretrain-lvd1689m
hf download briaai/RMBG-2.0 --local-dir models/BiRefNet/RMBG-2.0
git clone https://github.com/valeoai/NAF.git models/NAF
curl -L -o models/NAF/naf_release.pth https://github.com/valeoai/NAF/releases/download/model/naf_release.pth

The README's tested target is a 24 GB GPU with low_vram=true at 1024; people have squeezed it onto 8 GB with the 1024 cascade, but don't count on it.

Common issues

• First run feels frozen. It loads the full pipeline plus five aux models - give it a minute or two before judging.
• Instantly fails on the loader with "not found" - a model folder is missing or in the wrong place; the error names the exact path.
• OOM at 1536 - drop to 1024, lower max_num_tokens, or turn on low_vram at the loader.
• Bad background cutout - supply your own mask instead of fighting the auto-segmentation.

The workflow JSON in examples/ wires LoadImage → Load Model → this node → Save GLB. It ships with pai.jpg; drop your own image in and go.