Pixal3D CLIP Vision DINOv3 (CRT AutoDL)

Pixal3D is a picky eater: it doesn't read your image from a plain VAE or a generic CLIP. It needs a specific vision encoder - DINOv3 ViT-L/16 with a NAF (neighborhood attention) refinement head - because that's what makes its pixel-aligned trick work. The whole premise of Pixal3D is that it back-projects pixel features into 3D rather than injecting fuzzy image features through cross-attention, and that pipeline is only as good as the features it lifts. This node is what gets you that encoder.

CRTAutoDLPixal3DCLIPVision downloads dino_v3_L_naf_fp32.safetensors into ComfyUI/models/clip_vision/ on first run and loads it. There are no inputs at all - it's the "fixed file, download it for me, load it correctly" pattern every CRT AutoDL node follows. No model dropdown, no API key, nothing to tune.

The one output

CLIP_VISION - this is the clip_vision_model input of ComfyUI's native Pixal3DConditioning node. That node is where your preprocessed image, the DINOv3 features, and the camera FOV all get fused into the positive/negative conditioning that drives the shape and texture stages downstream. Without the CLIP vision wired in, the 3D pipeline has nothing to lift pixel features from.

Install

Same as the rest of the pack:

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

Restart ComfyUI. Or just use ComfyUI Manager and search CRT-Nodes. You need a ComfyUI recent enough to have the native Pixal3D/Trellis2 nodes (check for Pixal3DConditioning in the node search - if it's missing, update ComfyUI, not this pack).

Gotchas

• First run downloads ~a few hundred MB to a gigabyte. Watch the console for the CRT AutoDL progress bar. It's one-time; after that the file sits in models/clip_vision/ and any ComfyUI node can reuse it.
• It's fp32, deliberately. Pixal3D's NAF head expects the encoder at fp32 - that's why the file is bigger than the fp16 vision encoders you might be used to. Don't go "optimizing" it to fp16; you'll break the very feature fidelity this pipeline exists for.
• Wiring it into a non-Pixal3D workflow is possible but pointless. This is a specialized encoder tied to Pixal3D's conditioning. If you're doing plain CLIP vision work (IPAdapter-style), use the normal loaders instead.

This is a thin node - a one-click downloader and loader for a single fixed file - so there's not much else to say. The reason it exists as a separate CRT node at all is that the Pixal3D workflow needs six or seven moving parts, and the pack's philosophy is one node per part so you can assemble (and re-assemble) the graph without hunting model files. If you're building a Pixal3D workflow, this node is the piece that makes the difference between a graph that runs and a graph that dies on a missing encoder.