Pixal3D Model (CRT AutoDL)

This is the actual generator inside the Pixal3D workflow - the model that turns your image into 3D geometry instead of just another flat picture. Tencent ARC released Pixal3D in May 2026 as a direct challenge to TRELLIS.2 and Hunyuan3D: instead of loosely injecting image features through attention like older pipelines, it back-projects pixel features into 3D, so every pixel in your photo has a corresponding point in space. The community verdict, for what it's worth, was "looks like a step up from Trellis 2 raw." The whole thing is MIT-licensed, which is why Comfy Org repackaged it for ComfyUI and why this node can just hand it to you.

What the CRT AutoDL nodes do - this one especially - is remove the annoying part. You don't download anything, you don't check a model card to see which file goes in which folder. On your first run, CRTAutoDLPixal3DModel pulls pixal3d_bf16.safetensors (or the INT8 variant) into ComfyUI/models/diffusion_models/ with a progress bar in the console, and then loads it exactly like ComfyUI's own loader would. It's an auto-download node, not a wrapper around some cloud API - there's no key, no service, nothing leaves your machine.

The inputs that matter

• model_name - BF16 (default) or INT8 ConvRot. BF16 is the full-quality weights; INT8 ConvRot is the quantized variant for when the BF16 file won't fit. If you have a 24GB card, just leave it on BF16. If you're on 12–16GB and Pixal3D is already the heaviest thing in your graph, the INT8 conv-rotation quant is the difference between running and OOM-ing, and it looks close.
• attention_method - defaults to disabled, and that's the right call. The sageattn options (sageattn_qk_int8_pv_fp16_cuda, sageattn3, etc.) only do anything if you've installed the sageattention package yourself; without it the node just uses whatever ComfyUI has. pytorch attention is a safe manual fallback. Don't go hunting for speed here until the model works at all.
• patch_cublaslinear - off by default; it toggles ComfyUI's cublas_ops fast path. Leave it off unless you're debugging a specific performance issue.
• enable_fp16_accumulation - on by default, which is fine.

The one output

MODEL - feed this into the native Trellis2ShapeStage node. Pixal3D in ComfyUI is a two-stage pipeline (shape, then texture), and this is the model that drives the shape stage, with the DINOv3 CLIP vision feeding conditioning and the shape VAE decoding the result.

Install

ComfyUI Manager, search CRT-Nodes, install, restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/PGCRT/CRT-Nodes
pip install -r requirements.txt   # in your ComfyUI Python env

Then restart ComfyUI. First run downloads a several-GB model - that's expected, not a hang. Two gotchas specific to this node: it needs a recent ComfyUI, because it delegates to the core Trellis2/Pixal3D nodes in comfy_extras - if you get an import error, update ComfyUI before you blame the pack. And after a CRT-Nodes update, nodes that changed sockets can show up red with NaN values; right-click → Fix node (recreate).

Troubleshooting

• First run looks frozen - it's downloading. Watch the console for the progress bar; the Pixal3D BF16 file is a couple of gigabytes.
• OOM on BF16 - switch to INT8 ConvRot. The download is separate, so you'll grab the int8 file once and it stays on disk.
• "Trellis2ShapeStage not found" - your ComfyUI predates Pixal3D support. Update ComfyUI itself.
• Unwanted redownload - the node checks the target path before downloading, so if pixal3d_bf16.safetensors already sits in models/diffusion_models/, it just uses it. You can even drop the file there manually and skip the download entirely.

The rest of the pack's AutoDL family (CLIP vision, the two VAEs, MoGe, BiRefNet) is deliberately separate, one node per piece. Annoying on paper, nice in practice: you can reuse each downloaded file with ComfyUI's own loaders, and you swap pieces without re-fetching anything.