Loader ckpt+lora (Arrssenne)

This is the simpler half of the pack's loader pair. Where Loader ckpt+clip+vae+lora insists on loading CLIP and VAE from external files, this one says "the checkpoint has its own, use those" - so it has exactly two inputs (checkpoint and prompt) and still gives you five outputs. Fewer menus, fewer ways to misconfigure, and it's drop-in interchangeable with its big sibling.

When to use it

Checkpoints that embed a usable CLIP and VAE - the overwhelming majority of standard ones: SD1.5, SDXL, Pony, Illustrious, most merged .safetensors files. If your checkpoint is self-contained, this is the one you want. It loads the MODEL, CLIP, and VAE all out of the checkpoint file itself (output_clip/output_vae on), so there's no separate text encoder to pick and no clip_type to get wrong.

The pair is designed as a two-way fork: same five outputs on both, so you can swap this node in and out of a workflow with Loader ckpt+clip+vae+lora without touching anything else. Use the external-CLIP version only when the checkpoint can't supply its own text encoder (the author's example is the Krea2-style checkpoints that lack a usable one) or when you want a different encoder deliberately.

How it works

Exactly the same LoRA handling as the big loader: it parses <lora:name[:strength_model[:strength_clip]]> tags out of your prompt, matches them against installed LoRA files (exact path → path without extension → bare filename), applies each through the builtin LoraLoader so caching works, and returns the cleaned text. Unmatched tags are logged and skipped - never fatal.

The one real difference, besides the missing menus, is a nice failure mode. If the checkpoint genuinely has no usable embedded CLIP, this node raises a clear error instead of silently doing the wrong thing:

le checkpoint '...' ne contient pas de CLIP utilisable - utiliser 'Loader ckpt+clip+vae+lora (Arrssenne)' avec un CLIP externe

That's French for "this checkpoint has no usable CLIP - use the other loader with an external CLIP." It's the pack telling you exactly what to do, which is more than most loaders bother with. Heads up for non-French speakers: this pack's messages are in French; that's the one you'll actually see.

Inputs and outputs

• ckpt_name - dropdown of your checkpoints.
• text - forceInput; wire your prompt (with any <lora:...> tags).

Outputs: MODEL, CLIP, VAE, text (tags stripped - wire this into CLIP Text Encode, not your raw prompt), and checkpoint_name (the raw string, for Filename ckpt+seed / Image Saver metadata). Identical signature to the other loader, which is the point.

Install

ComfyUI Manager → search "ComfyUI-Arrssenne", or:

cd ComfyUI/custom_nodes
git clone https://github.com/arrssenne/comfyui-arrssenne

Restart, find it under Arrssenne/Loaders. No models, no keys, no heavy Python deps - the pack's only real requirement is Pillow (already in every ComfyUI), and this node doesn't even use it.

Where does this land on the "reach for it" scale? If you've ever loaded a workflow where someone typed CLIP type: ltxv on an SDXL loader and watched the encode explode, you'll appreciate a loader that can't make that mistake. It's the everyday option of the pair: fewer knobs, no misconfiguration path, and a friendly error when it genuinely can't do the job.