Load LVCD Model

If you've ever wanted to feed a lineart animation a single color reference frame and get a whole consistently-colored video back, you've landed on the right pack. LVCD is a SIGGRAPH Asia 2024 model from luckyhzt for reference-based lineart video colorization, and this node is the front door to it. It's not a loader in the "pick a file" sense - it's the node that assembles the entire two-model pipeline and hands you the LVCD_pipe that everything else in this pack consumes. No LVCD_pipe, no video.

The name is a slight lie in a good way: it doesn't just load the LVCD adapter. It takes a plain Stable Video Diffusion checkpoint from your models/checkpoints folder and straps LVCD's sketch-guided ControlNet onto it, building the full SVD + adapter stack in one shot. This is the classic Kijai move - the wrapper author behind KJNodes, WanVideoWrapper and the LivePortrait port, who has a reputation for shipping the plumbing so you don't have to think about it. You load two models with one node and never see the Config instantiation.

What you actually set

The node has only three controls, and two of them are "leave it alone":

• model - a dropdown of everything in ComfyUI/models/checkpoints. The author's own tooltip says it plainly: "Normal SVD model, default is the normal very first non-XT SVD." The original stable-video-diffusion-img2vid works best, though the README notes 1.1 and XT run too. Kijai's pruned fp16 svd-fp16.safetensors from Kijai/LVCD-pruned is the lighter option and what the bundled example workflow uses.
• precision - fp16 (default), fp32, or bf16. fp16 is right for almost everyone; fp32 is the "something looks wrong, check whether it's precision" escape hatch that roughly doubles VRAM; bf16 is the better pick if you're on an Ampere-or-newer card and your setup plays nice with it.
• use_xformers - off by default, and modern ComfyUI builds usually don't need it. Flip it on only if you're on an older install where xformers is your actual attention path.

The output is a single LVCD_pipe socket, and that's the whole contract. It wires straight into both LVCDSampler and LVCDDecoder.

How it works under the hood

Behind the scenes the node is doing real work: it checks whether the LVCD adapter exists at ComfyUI/models/lvcd/lvcd-fp16.safetensors and if not, auto-downloads it from Kijai/LVCD-pruned via huggingface_hub. Then it loads the SVD config with the LVCD ControlNet injected, and swaps the attention implementations based on your use_xformers flag. That's the "handshake" - SVD supplies the temporal prior and the image-conditioning stack, LVCD's ReferenceAttention and masked temporal attention supply the color transfer from your reference frame. You never touch either model directly again.

Installing it

The real install is the pack-level one, so this is the same for all three nodes here:

cd ComfyUI/custom_nodes
git clone https://github.com/kijai/ComfyUI-LVCDWrapper

or just search "LVCD" in ComfyUI Manager and let it handle the rest, then restart ComfyUI. Two model pieces matter: an SVD checkpoint in models/checkpoints, and the LVCD adapter, which lands in models/lvcd automatically the first time you run the node. This pack's requirements.txt is heavier than most - omegaconf, open-clip-torch, pytorch-lightning, timm, plus clip straight from the OpenAI GitHub. If Manager complains or the node errors on import, make sure those are actually installed in the environment you run ComfyUI with.

Where people get burned

• Empty checkpoint dropdown. You don't have an SVD in models/checkpoints. The original is ~2GB from Stability; the pruned fp16 is a little under that. This node will not work with an SD 1.5 or SDXL checkpoint - it has to be SVD.
• The auto-download silently fails. The LVCD adapter fetch needs huggingface_hub to reach the internet. If you're offline or behind a proxy, drop lvcd-fp16.safetensors into models/lvcd manually.
• VRAM. This is SVD under the hood - a real video diffusion model. Expect it to be hungry; fp16 at 576×576 is comfortable on 12GB+ and a squeeze below that. The sampler's offloading behavior (see the LVCD Sampler article) is your main lever.

It's a niche pipeline - by 2025-2026 most people reach for Wan or LTX for video instead - but if the job is "colorize this lineart with this reference," this is still the honest, purpose-built way to do it, and it all starts with this node.