AniDocLoader

AniDocLoader is where every ComfyUI-AniDoc workflow starts, and it's the most unusual loader you'll meet this week: it takes no inputs at all. You drop it on the graph, it loads the whole AniDoc pipeline, and it hands you an ANIDOC_PIPELINE socket that feeds straight into AniDocSampler. One per workflow, nothing to configure, and - here's the catch - the first run downloads the models.

What it's loading

AniDoc (arXiv 2412.14173) colorizes line-art animation video from a single colored reference. To do that it doesn't train a whole new video model; it fine-tunes an existing one. So this loader is really loading two big things:

1. Stable Video Diffusion Img2Vid (SVD XT 1.1) - the base video diffusion model, in fp16. This is the actual video backbone, and it's why the pack wants a serious GPU.
2. The AniDoc UNet + ControlNet - the fine-tuned pieces (Yhmeng1106/anidoc) that make SVD follow your lineart condition and your reference's colors.

In code terms: it builds a UNetSpatioTemporalConditionModel plus a ControlNetSVDModel, wraps them in a custom AniDocPipeline (an SVD img2vid variant), casts to fp16, and enables xformers memory-efficient attention if your ComfyUI build has it. The result is cached on the node, so reruns are near-instant; the cost is only the first load.

The first-run download

When you first execute the node it calls snapshot_download on two HuggingFace repos and dumps them into ComfyUI/models/diffusers/:

• Yhmeng1106/anidoc → models/diffusers/anidoc
• vdo/stable-video-diffusion-img2vid-xt-1-1 (fp16 config + weights, no UNet since AniDoc brings its own) → models/diffusers/stable-video-diffusion-img2vid-xt

That's several gigabytes and it happens in the middle of your first queue, so don't panic when the progress bar stalls on the very first node. Let it finish. If the auto-download fails (flakey connection, HF hiccup), python install.py in the custom node folder does the same download outside ComfyUI, and the README has manual download links for both repos.

One genuinely annoying mismatch: the README says SVD lives in stable-video-diffusion-img2vid-xt-1-1, but the code looks in stable-video-diffusion-img2vid-xt (no -1-1 suffix). The code wins. If you manually download the SVD files, put them where the loader actually looks, or you'll get a from_pretrained error and a mystified afternoon.

Install and first steps

cd ComfyUI/custom_nodes
git clone https://github.com/LucipherDev/ComfyUI-AniDoc
cd ComfyUI-AniDoc
python install.py   # requirements + model downloads + CoTracker package

Or search "ComfyUI-AniDoc" in ComfyUI Manager. Either way, restart ComfyUI afterward. The nodes live under the AniDoc category.

Gotchas

• VRAM is the real constraint. This is an SVD-class pipeline in fp16 - you want a 12GB+ card to breathe, and you'll be glad the sampler has a decode_chunk_size knob for when you don't.
• "Couldn't download models" is non-fatal - the loader logs it and continues; the error that follows will tell you exactly which file is missing.
• You still need CoTracker's checkpoint manually for the tracking path (cotracker2.pth into models/cotracker), and VideoHelperSuite is required by the example workflow that loads frames and saves the result. The nodes themselves don't import VHS - it's the workflow around them that needs it.

Honest framing: this is a wrapper around a research pipeline, not a polished product. It works, and it's genuinely the easiest way to run AniDoc locally. Just know the first click is the expensive one.