Anymatix Checkpoint Loader

At first glance AnymatixCheckpointLoader looks like ComfyUI's built-in CheckpointLoaderSimple wearing a different name. Same three outputs - MODEL, CLIP, VAE - same job: load a diffusion checkpoint so the sampler can denoise latents with it, the CLIP can encode your prompt, and the VAE can round-trip pixels to and from latent space. The difference is the one input. The stock loader gives you a dropdown of whatever checkpoints happen to be in models/checkpoints on your machine. This one takes a plain text string that is treated as a file path.

Why would you want that? Because the whole Anymatix pack is built for workflows that are written once and run somewhere else. A card in the Anymatix app references a checkpoint by URL; AnymatixFetcher downloads it to whatever machine is running the graph and hands this loader the resulting path. The workflow JSON never had to guess what was in your folder - at the moment the graph runs, the file exists, and the path is already correct.

That design also means the node is genuinely useful in stock ComfyUI, not just inside the app. Wire AnymatixFetcher into it and you get a "load this checkpoint from this URL" workflow that works on any machine with an internet connection. Or type an absolute path to a checkpoint that lives somewhere models/checkpoints doesn't cover.

How it works

Under the hood it's a thin, honest wrapper. It verifies the file exists (and raises a clear FileNotFoundError if it doesn't, naming the path), then calls ComfyUI's own load_checkpoint_guess_config with your path - the same guess-the-architecture loader the built-in node uses, so SD 1.5, SDXL, and the newer diffusion transformers are all handled automatically. The first three of its outputs - MODEL, CLIP, VAE - are what you wire onward.

A trap worth knowing: unlike some twins in this pack, this loader does not search your models directories for a bare filename. It wants a real path. If you type just my_model.safetensors it will fail, because it's checking the literal string. Feed it the fetcher's output, an absolute path, or use one of the pack's basename-resolving loaders instead.

The inputs and outputs

• ckpt_name - the only required input. A STRING, intended to be the full path to the checkpoint (or AnymatixFetcher's STRING output).
• MODEL → the KSampler. CLIP → a positive/negative prompt pair. VAE → VAEEncode/VAEDecode.

Install

This is the shared pack story:

cd ComfyUI/custom_nodes
git clone https://github.com/Anymatix/anymatix-comfy-nodes

restart, or just hit "Install Missing Custom Nodes" in ComfyUI Manager (search "anymatix-comfy-nodes"). No extra model downloads on top of the checkpoint itself - the node is a loader, not a fetcher, so a checkpoint you don't have yet will fail until AnymatixFetcher (or you) puts it on disk.

Issues to expect

The most common beginner stumble is mixing up the two families of loaders in this pack. Several sibling nodes (the CLIP and LoRA-model-only twins) do resolve bare filenames against your model folders; this one takes the path literally. If you're hand-typing, use the full path. And remember a checkpoint load on a low-VRAM card is the usual suspect for OOM at the start of a run - that's the model, not this node, but knowing it's a loader helps you reason about where the memory went.