☁️BizyAir CLIP Set Last Layer

This is the "CLIP skip" node, just running against BizyAir's cloud-loaded CLIP instead of a model on your own disk. If you've used A1111's "Clip skip" slider or ComfyUI's regular Clip Set Last Layer node, this behaves identically - it just needs a BIZYAIR_CLIP input, not a plain CLIP, because BizyAir's nodes only talk to each other's cloud handles.

Why this exists

CLIP skip conditions your prompt on an earlier internal layer of the text encoder instead of the very last one, which reads the prompt through a slightly more general representation. It's not a universal quality knob - it's a compatibility convention that traces back to NovelAI, which conditioned on CLIP's second-to-last layer back in 2022. Every anime-lineage checkpoint since (Illustrious, Pony, NoobAI, and their SD 1.5 ancestors) inherited that expectation, and their release notes still print "Clip Skip: 2" as a result.

There's a real gotcha worth knowing before you touch the slider: on SDXL and everything derived from it, the model already defaults to the penultimate layer. Setting it to -2 on an SDXL-family checkpoint mostly restates a default you already have. What actually breaks things is setting it to -1 on a Pony/Illustrious model - that's the mistake, not the fix. CLIP skip is a genuinely meaningful dial on SD 1.5; on SDXL derivatives it's closer to "don't set it to the wrong thing."

The one input that matters

stop_at_clip_layer - an integer from -1 (default, no skip) down to -24. In practice you'll almost never go past -2: -1 is correct for realistic SD 1.5 checkpoints, -2 is the standard for anime-lineage SD 1.5/SDXL checkpoints, and anything deeper than that is unusual territory most models were never trained to expect. The slider's -24 floor is generous rather than a real recommendation - it's there because BizyAir's node doesn't clamp per-model, not because most CLIP encoders have that many usable layers.

clip in, patched BIZYAIR_CLIP out - wire it between your BizyAir checkpoint/CLIP loader and whatever encodes your prompt downstream.

Installing it

ComfyUI Manager: search BizyAir, install, restart. Or by hand:

cd ComfyUI/custom_nodes && git clone https://github.com/siliconflow/BizyAir.git

Restart after. No model files to fetch for this specific node - it operates on a CLIP that's already loaded cloud-side by an earlier BizyAir node in your graph. You do need a BizyAir account and API key set up before any node in this pack will actually execute.

Where people get burned

Feeding it a plain CLIP instead of BIZYAIR_CLIP. This node's input and output types are both BizyAir-specific, so it only chains with other BizyAir loader and conditioning nodes - you can't drop it into a workflow built around ComfyUI's stock checkpoint loader.

Assuming -2 always "does something." On an SDXL-family model it's frequently a no-op relative to the model's own default, so if you're not seeing a visible difference, that's expected, not a bug in the node.

Applying it to a 2026-era model. Newer architectures moved from CLIP to LLM-class text encoders (Qwen3, Mistral) that have no CLIP layers to skip at all - the setting is meaningless there regardless of what value you set.