Load VCG Model

Every VideoColorGrading workflow in this pack starts at the same place, and it's this node. Load VCG Model reads a single combined checkpoint and hands you a vcg_pipeline that the color-grading nodes are helpless without. It's the most boring node in the pack - one dropdown, one output - but if you get the model file in the wrong folder, it's also the first thing that breaks. So let's get it right.

What it loads

That one checkpoint actually contains four models, packed together with key prefixes:

• CLIP ViT-B/32 - the image encoder used to match your reference frame to your source footage.
• VAE - encodes the reference and matched frame into SD1.5 latents.
• ReferenceNet - an SD1.5-architecture UNet that extracts the "what's different between these looks" attention features.
• L-Diffuser - the second UNet that denoises the LUT itself.

The node pulls them apart, builds each with the right ComfyUI machinery, and sends them out as one vcg_pipeline object. You never see the parts; you just get one wire. All four are needed by VCGGenerateLUT, which is the only thing this output plugs into.

The input

Just one: model_name, a dropdown listing checkpoints from your ComfyUI/models/checkpoints folder. That folder is the gotcha - people instinctively drop the file in models/vae or models/unet, but this node reads the standard checkpoints directory, same as any SD checkpoint loader.

The output

vcg_pipeline (VCG_PIPELINE type) → straight into VCGGenerateLUT. That's the entire fan-out. The loader does no inference itself; nothing visual happens until the generate node runs.

Install and the model file

Same story as the rest of the pack:

• ComfyUI Manager: search "VideoColorGrading" and install, or
• cd ComfyUI/custom_nodes && git clone https://github.com/kijai/ComfyUI-VideoColorGrading then restart.

No requirements.txt, no extra pip packages - it uses only ComfyUI's own stack. But it's written against ComfyUI's newer node API (comfy_api.latest), so if the node fails to register, update ComfyUI first. And the pack was designed for ComfyUI Manager, so that's the friction-free route.

Then download the model from Kijai's HuggingFace repo: vcg_combined_fp16.safetensors, about 4.1GB, Apache-2.0 licensed, into ComfyUI/models/checkpoints. Restart (or just refresh) and it'll show up in the model_name dropdown.

Why trust the source

If the "kijai" in the repo name means nothing to you yet: this is the author who ships ComfyUI support for essentially every major video model within days of release - Wan, LTX, SUPIR before that - and who's known for dropping one-line fixes that make launch-week models actually runnable on normal GPUs. This pack is a port of the ICCV 2025 paper Video Color Grading via Look-Up Table Generation, and it's the standard pattern for his ports: one combined checkpoint, minimal nodes, no build steps. If the workflow you found uses these nodes, the only real prerequisite is this one 4.1GB file in the right folder. Everything else is vanilla ComfyUI.

Troubleshooting

• Model name dropdown is empty - the .safetensors isn't in models/checkpoints. Move it there.
• Missing node / "not found" on load - stale ComfyUI; update it, since the pack targets the newer extension API.
• Out of memory during generation - that's the VCGGenerateLUT node's job to fix (it's the VRAM-heavy part), not this loader. The checkpoint itself is ~4.1GB, so loading it on a 12GB card is a non-event.