Bubba Combo Loader

The Bubba Checkpoint Loader is the plain one - checkpoint in, model/CLIP/VAE out, done. Bubba Combo Loader is the one you reach for when you're picky, because it loads the checkpoint and lets you override the VAE, swap the CLIP/text encoder, and set a CLIP skip, all in a single node. That's normally three or four nodes in a stock ComfyUI graph - Load Checkpoint, a VAE Loader, a separate CLIP Loader, and a CLIP Set Last Layer - collapsed into one.

When picky matters

A surprising amount of the time, actually. Most anime SDXL checkpoints ship with the baseline SDXL VAE, and people routinely swap in a known-good external VAE because the bundled one can produce washed-out or garbled colors. With this node you just pick vae_name from a dropdown and it replaces the checkpoint's VAE - no extra loader, no rewiring the graph. Same for the text encoder: if you run a model that pairs better with a specific CLIP (or a newer text encoder), clip_name handles it, with clip_type to say which kind you're loading.

clip_skip is the one that's purely about flavor. On SDXL-lineage anime models - Illustrious, NoobAI, Pony - skipping the last CLIP layer is a genuinely popular setting that shifts the aesthetic: people set it to 1 or 2 to change how the model interprets tags, and with this node it's an integer on the same panel as everything else instead of a separate node you'd forget to wire. It's recorded in the pipe's metadata, so the choice doesn't get lost either.

The inputs, sorted by what you'll actually touch

• ckpt_name - required, the checkpoint dropdown.
• clip_name and vae_name - required dropdowns, but they default to the checkpoint's bundled versions, so you only change them when you want an override. If you leave them at their defaults you're effectively getting the plain loader's behavior.
• clip_skip - 0 to 24, default 0 (no skip). Set 1 or 2 on anime SDXL models and A/B it.
• clip_type - tells the loader what kind of CLIP file clip_name is; the default matches the common case.
• clip_device - optional, default unless you have a reason to pin CLIP to a specific device (e.g. forcing it off the main GPU to save VRAM).

Everything else is the pipe contract: optional pipe in, updated pipe + metadata out, plus model/clip/vae for people who don't want to use the pipe, and a checkpoint_name string output.

Where it fits

This is the recommended starting point in the pack's own quick-start: Combo Loader (or Checkpoint Loader) to create a pipe with the model stack, then LoRAs, prompt building, sampling, and saving all ride the pipe. If your workflow is "one checkpoint, fine as-is," the plain loader is less to think about. If your workflow is "this checkpoint plus my favorite VAE plus clip skip 2, every time," the Combo Loader saves you a couple of nodes and, more importantly, keeps the settings attached to one place where you can see them at a glance.

Install and gotchas

Bubba Nodes is one pack, 37 nodes, installed like any other custom node:

cd ComfyUI/custom_nodes
git clone https://github.com/bubbafett5611/Bubba_Nodes

or ComfyUI Manager → search "Bubba Nodes". Restart after. The pack requires ComfyUI 0.27.0+ (it uses the newer V3 registration API) and pulls aiohttp + pydantic from its requirements.

Where people get burned: clip skip is not a free slider on every model. It does nothing useful on models that were never trained around it, and on some instruct-tuned encoders the "CLIP skip" concept barely applies. Change it, A/B it, and if the images are indistinguishable or worse, set it back to 0 - the metadata will happily record whatever you picked, but the model decides whether it matters.