Chord - Load Model

ChordLoadModel is the least glamorous node in the ComfyUI-Chord pack and the one everyone trips over first, because it has a prerequisite that's easy to miss: a big checkpoint file you have to fetch by hand, from behind a gate. Get past that and this node is genuinely boring - which is the point. It loads, you wire it, you forget it.

What it actually loads

Look at its one input, ckpt_name (a COMBO), and you'll see the dropdown lists your ComfyUI/models/checkpoints folder - the same folder where your SD/SDXL checkpoints live. That's the tell: Chord is not a small estimator you slot into a side channel. It's a full diffusion model built on an SD 2.1 base (fp16, with a UNet, VAE and text encoder all packed in). That's why the file is several gigabytes, why it lives in checkpoints, and why it eats GPU memory like a real checkpoint rather than a lightweight upscaler.

The node hands you one output, chord_model (a custom chord_model type). Wire it straight into the chord_model input on ChordMaterialEstimation and nothing else consumes it.

Mechanically there's some quiet engineering worth knowing. It loads the config from config/chord.yaml, then reads the safetensors with ComfyUI's safe loader. The interesting bit: it remaps the CLIP text-encoder state dict keys on the fly, so the checkpoint keeps working even after transformers jumped versions and shuffled its internal layout. Old pack, new library - handled. The model is wrapped in a ModelPatcher, which is how ComfyUI manages GPU/CPU offloading, so it coexists with your other models without needing everything resident at once.

Getting the model file (the part that bites)

The README is blunt: download chord_v1.safetensors from the Hugging Face repo and drop it in ./ComfyUI/models/checkpoints. The bite: the repo is gated. You need a Hugging Face account, and you have to agree to share your contact information to access the file. No token in the ComfyUI settings this time - just download it through your browser, agree to the terms, and place the file.

If the dropdown doesn't show chord_v1.safetensors, it's almost always one of two things: the file landed in the wrong folder, or ComfyUI hasn't refreshed its checkpoint list - restart, or look for the refresh in the model lists.

Install and gotchas

• ComfyUI Manager → search "ComfyUI-Chord" → install, then restart.
• Or cd ComfyUI/custom_nodes && git clone https://github.com/ubisoft/ComfyUI-Chord.git and pip install -r ComfyUI-Chord/requirements.txt (diffusers, omegaconf, imageio).

Now the part the community will tell you before the README does: the model is under the Ubisoft Machine Learning License (Research-Only - Copyleft). That is not "open source," and it is not "just download it and ship your game." It's source-available, research and non-commercial use only, and the Reddit thread that first surfaced the pack spent most of its comments arguing exactly that. If you're making an asset pack to sell, read the license before you build on this.

One more failure mode, straight from the source code: if you point it at a checkpoint that isn't actually the Chord weights, it raises "Failed to load model, check if the checkpoint file is correct." That's not a cryptic error - it means you grabbed the wrong file or a corrupted download. Re-download from Hugging Face and try again.