ZeroClip-D Conditioning (SDXL)

SDXL support in this pack follows one consistent shape, and ZeroClip-D Conditioning (SDXL) is the D variant of it: two anchor libraries in - one for the sequence path (D=2048), one for the pooled path (D=1280) - one seed, and an SDXL-format CONDITIONING with pooled_output out. Same self-bootstrapped anchor idea as the base D node, rebuilt for SDXL's dual-encoder reality.

If you've already wired the SDXL versions of the A or B conditioners, this will feel like a rerun with a different load type. The inputs even carry the same dimension tooltips (anchors_seq = 2048, anchors_pooled = 1280), and the failure modes are the same two: not enough libraries, or the wrong dimensionality.

The inputs

• anchors_seq - bootstrap anchor library for the sequence embeddings (D=2048), from a ZeroClip-D Load Anchors node pointed at the SDXL seq bootstrap file.
• anchors_pooled - bootstrap anchor library for pooled embeddings (D=1280), from a second Load Anchors node pointed at the SDXL pooled bootstrap file.
• seed - the standard ZEROCLIP_SEED.

Output: SDXL-format conditioning. Use an SDXL-format Empty Conditioning (sdxl toggled on) for the negative socket.

How it works

Exactly the A-style algorithm, run twice: the same packed seed drives coherent-noise weighted sums over the seq library and the pooled library separately, each L2-normalized, then combined into the two-part SDXL conditioning dict. Because one seed drives both, the sequence and pooled concepts stay locked together - which is what SDXL needs to keep its two encoder paths coherent.

Install

Part of ComfyUI-ZeroCLIP-nodes. ComfyUI Manager (search "ZeroCLIP"), or:

cd ComfyUI/custom_nodes
git clone https://github.com/MushroomFleet/ComfyUI-ZeroCLIP-nodes

Restart ComfyUI; it's under ZeroClip/D - Self-Bootstrapped. No pip install. You need the SDXL bootstrap anchor files (seq + pooled) from huggingface.co/mushroomfleet/zeroclip in models/zeroclip/.

Where people get burned

• Both sockets are required. Wire only anchors_seq and the node errors on the missing anchors_pooled. This is the #1 mistake on every SDXL node in this pack - two loaders, always.
• The bootstrap files are the expensive ones to lose. The D build is a ~5-hour GPU job. If you delete or forget the SDXL bootstrap files, do not re-run the build out of impatience - just re-download from Hugging Face.
• D on SDXL is even harder to read than D on SD1.x. With two anchor libraries both derived from model probing, there's even less intuition about what a seed "means." If you're using SDXL and want predictable text-free control, the A SDXL nodes are the sane choice; keep D for the research experiments.