ScarlotSoft Checkpoint + CLIP Skip

Every anime workflow on SD 1.5 has the same two-node ritual: a CheckpointLoaderSimple and a CLIP Set Last Layer set to -2. ScarlotSoft Checkpoint + CLIP Skip collapses that into one loader. It loads the checkpoint the normal way - same ckpt_name dropdown you already know, backed by ComfyUI's own load_checkpoint_guess_config - and then applies your CLIP skip inside the loader before the CLIP ever leaves the node. Out of the box you get MODEL, CLIP, and VAE wires just like the core loader, plus a stop_at_clip_layer dial.

That dial is the whole reason this node exists. It runs from -1 to -24 with a default of -1 (no skip), and the value you mostly want is -2. The mechanics are the same CLIP skip that has been a Danbooru-lineage convention since NovelAI: stop the text encoder a layer early so it conditions on a slightly more general representation. For SD 1.5 anime checkpoints - and their descendants through Pony, Illustrious, and NoobAI - that's -2, printed right on the release notes. For realistic SD 1.5 it's usually -1. For SDXL, honestly, don't bother: both SDXL text encoders already default to the penultimate layer, so setting 2 usually changes nothing and setting 1 is what turns Pony into noise. And on 2026 LLM-encoder models there's no CLIP at all to skip, so this dial is meaningless there.

Where the node earns its keep is workflow cleanliness. Burying the skip in the loader means one fewer node and one fewer way to forget to set it. If you swap checkpoints between an anime model and a realistic one, you'll be flipping this dial anyway, so having it on the same card as the model picker is genuinely ergonomic. It's a thin wrapper - clone the CLIP, call clip_layer() - but thin wrappers around a setting you always touch are the good kind.

Install with the rest of the pack via ComfyUI Manager (search ScarlotSoft), or:

cd ComfyUI/custom_nodes
git clone https://github.com/scarlotsoft/ComfyUI_ScarlotSoft

then restart. No dependencies, no model downloads - this one is pure plumbing over ComfyUI's own loader. As with all ScarlotSoft nodes it wears the suite's dark theme; hard-refresh if it looks off after a UI-mode switch.

The trap to keep in mind isn't the node, it's the convention: if you're on an SDXL-based anime model and an image looks worse at -2 than at -1, you're not doing it wrong - SDXL just doesn't want the skip. And if you load a workflow built around this node with a missing-CLIP-skip panic, remember the image's PNG metadata holds the original graph, so drag your old output back in and see exactly what it was set to. One honest caveat: the pack has a tiny community footprint, so don't hunt for forum threads on this node - it does one thing, and you can read its source in nodes/scarlotsoft_checkpoint_clip.py if you ever doubt it.