[SK] Checkpoint

This is ComfyUI's "Load Checkpoint" node, unchanged in what it does, changed in how you find the file. Point it at a checkpoint and it hands you the three outputs everything downstream needs - MODEL, CLIP, VAE - exactly like the stock node. The only reason it exists is the picker: instead of one long alphabetical dropdown, ckpt renders as a folder tree, so if you keep your checkpoints organized into subfolders (by architecture, by NSFW/SFW, by whatever your own system is), you browse it the way you actually organized it instead of scrolling.

It's part of SK-ComfyUI-FolderingLoader, a small pack by GitHub author SKFRMSEHF that reimplements seven or eight of ComfyUI's own loaders with that same tree-picker treatment. It's genuinely obscure - no README content shipped, no discussion of it anywhere I could find - so treat it as a solo-dev convenience tool rather than something with a support community behind it. That's fine for what it is: the outputs are standard ComfyUI types, so it drops into any workflow that expects a normal checkpoint loader, and you can mix it freely with core nodes or other packs' loaders in the same graph.

How it works

A checkpoint file bundles three separate components into one archive: the diffusion model (MODEL), the text encoder (CLIP), and the latent codec (VAE). Loading it just means reading that file and splitting it back into the three pieces ComfyUI's graph treats as independent objects you can wire wherever they need to go - sampler, text encode, VAE decode. Nothing about that mechanism is special here; the node is doing exactly what CheckpointLoaderSimple does. The only added logic is building the folder-tree structure for the picker itself.

The input that matters

There's really only one: ckpt, the checkpoint file, picked from the folder tree your checkpoints/ directory produces. That's it - no config file, no extra toggles. (If you need the old-style loader that also takes an explicit .yaml config, this pack ships that separately as SK_CheckpointLoader, marked deprecated by its own display name - you almost certainly don't need it.)

Outputs are MODEL, CLIP, and VAE, in that order, matching what every downstream node in a standard workflow expects.

Installing it

No install instructions came with the pack - the README is empty - so this is the standard route: ComfyUI Manager → search SK-ComfyUI-FolderingLoader (or the class name) → install → restart. Or manually: cd ComfyUI/custom_nodes && git clone https://github.com/SKFRMSEHF/comfyui_SK_Loader, then restart ComfyUI. It downloads no models of its own - you bring the checkpoints, same as with the core loader.

Common issues

• Empty tree, no files listed. The picker only shows what's already in your checkpoints/ folder (or wherever your extra_model_paths.yaml points). Drop the file in and restart - this isn't a live filesystem watcher, it builds the list at startup.
• .ckpt vs .safetensors. If your combo shows both formats for the same model, take the safetensors one. .ckpt uses Python's pickle format under the hood, which can execute arbitrary code on load; safetensors stores tensor data only. There's no quality difference between them, only a security one.
• Full checkpoint vs pruned. If you accidentally grabbed a "full" training checkpoint (with optimizer state baked in) instead of the pruned fp16 release version, it'll load fine here but eat roughly double the disk and RAM for identical output - nothing this node can warn you about, since it just loads whatever file you point it at.
• No community thread to fall back on. This pack has no visible discussion anywhere, so debugging is you and the GitHub repo. If it ever breaks after a ComfyUI update, the stock CheckpointLoaderSimple is a drop-in functional replacement while you wait for a fix.