[Inference.Core] Anime Lineart

Run a photograph through a generic line extractor and you get realistic, slightly messy contour lines. Run an anime illustration through the same tool and it often looks wrong - too much texture noise, lines that don't match the clean, deliberate strokes anime art actually uses. Anime Lineart is trained specifically on illustration-style linework, so it extracts the kind of clean, confident lines an anime artist would actually draw, rather than every gradient a photo-trained model happens to notice. It's the preprocessor of choice for keeping a character's silhouette and linework consistent across generations - extract lines from a reference, then regenerate with a different style or coloring while the composition stays locked.

How it works

This is a dedicated line-extraction network trained on anime-style source images rather than photographs, so its notion of "what counts as a line" matches how illustrations are actually constructed - clean outlines and deliberate interior strokes - instead of picking up every texture gradient a general-purpose edge detector would flag on a photo.

The inputs and outputs that matter

Minimal by design:

• image - required, your source illustration.
• resolution (default 512) - working resolution.

One output: IMAGE, the extracted line-art map, wired into a lineart ControlNet (specifically the anime-lineart-trained variant, not the generic/realistic one - see troubleshooting).

How to install it

ComfyUI Manager: search ComfyUI-Inference-Core-Nodes, install, restart. By hand:

cd ComfyUI/custom_nodes
git clone https://github.com/LykosAI/ComfyUI-Inference-Core-Nodes

then run install.py, or pip install -e .[cuda12] (or .[cuda] / .[rocm] / .[directml] / .[cpu]). Restart ComfyUI. The model weights download from Hugging Face the first time you run the node.

Common issues & troubleshooting

The single mismatch that trips people up: this node is trained for anime-style input, and feeding it a photograph gives noticeably weaker, less confident lines than this pack's other Realistic Lineart node would. Match the preprocessor to your source style - anime art gets Anime Lineart, photos get Realistic Lineart.

The second one is on the ControlNet side, not this node: most lineart ControlNet checkpoints are trained on a specific lineart style (anime vs. realistic), so double-check you're loading the anime-lineart variant to match this preprocessor's output. Mixing a realistic-lineart ControlNet with this node's anime-style output tends to produce results that technically follow the lines but fight the prompt in strange ways. And if your source image is a scanned manga page or a screenshot with halftone dots or compression noise rather than clean digital art, this pack's Manga Lineart node (built specifically to denoise that kind of source first) will usually give you a cleaner result than this one.