Bubba Model Components Override

Bubba Model Components Override is the surgical tool for the "I like this checkpoint, but its bundled CLIP/VAE is wrong for what I'm doing" problem. It takes a Bubba pipe and selectively replaces two of its components - the CLIP/text encoder and the VAE - then applies CLIP skip, without touching the model. Drop in an external CLIP or a better VAE, leave everything else in the pipe alone, and keep flowing. Selections left at None preserve the corresponding pipe component, which is the whole design.

Why would you want that? Two very common real-world cases. First, CLIP skip: some tag-based checkpoints want a couple of CLIP layers skipped to reduce the mid-layer embeddings the community finds harmful - and the override is the node that applies it on a pipe. Second, the model-comparison rig: the README's comparison workflow explicitly uses this node to attach a shared external CLIP or VAE before Bubba Model Compare Loader, so every checkpoint under test uses the same components instead of each one's bundled defaults - which makes the comparison actually about the model weights, not the incidental text encoder or VAE differences.

How it works

The inputs are pipe plus four selectors: clip_name (a dropdown of available CLIP/text encoder files, default "None (keep pipe component)"), vae_name (same pattern for VAEs), clip_type (the encoder flavor, stable_diffusion by default - set it to match the CLIP file you're loading, e.g. for different encoder architectures), and clip_skip (0 to 24, default 0 - the number of CLIP layers to skip during encoding). clip_device defaults to default and lets you pin where the clip runs if you're managing device placement.

Any selection left at "None (keep pipe component)" leaves the pipe's existing value untouched. The outputs are the updated pipe and metadata, plus direct model, clip, and vae outputs - so even outside the pipe system you can use this node as a convenient "load a clip with CLIP skip baked in" helper. info summarizes what changed.

The inputs that matter

clip_name, vae_name, and clip_skip. That's the whole job. If you only need a CLIP-skip'd encoder for a pipe workflow, this node replaces the older habit of a separate CLIP loader + a text-encode wrapper.

How to install it

It's a Bubba_Nodes pack node, so install the pack once. ComfyUI Manager → "Bubba Nodes" → install → restart, or:

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

restart ComfyUI, and pip install -r requirements.txt if needed (aiohttp, pydantic - light). Requires ComfyUI v0.27.0+ for the V3 node API.

Gotchas

The mental model to keep straight: this node only works on pipes that already carry components. Feed it a partial pipe with no CLIP and leave clip_name at "None (keep pipe component)" and you'll end up with no CLIP - "None" means keep, not auto-fill. In the model-comparison context, remember the README's pairing rule: if you attach an external CLIP/VAE here, disable the Compare Loader's replace_clip/replace_vae toggles, or the loader will overwrite your carefully prepared components with each checkpoint's bundled ones. And a quick conceptual note from the KB's concepts essay: CLIP skip is a property of the CLIP-encoder family - on modern LLM-encoded models (Flux-class, Anima) there's no CLIP to skip, so this node is most at home on the SDXL/tag-based lineage it was clearly built around. The pack is new with no community discussion yet; the README's model-comparison section is the best guide to wiring this into a multi-branch compare.