☁️BizyAir MinusZone ChatGLM3 Text Encode

Most ComfyUI text-encode nodes assume CLIP, or T5 on the newer flow-matching architectures. This one doesn't, because the model it's built for doesn't use either. Kolors, the text-to-image model Kuaishou (Kwai) released in mid-2024, made the unusual architectural choice of using ChatGLM3, a Chinese-developed large language model, as its text encoder instead. That choice is exactly why this node exists as its own thing rather than being folded into a generic BizyAir CLIP encoder: ChatGLM3 needs its own tokenizer and encoding path, and the output it produces isn't a standard CLIP-typed conditioning object.

The README credits where this integration actually came from: BizyAir's Kolors support, this node included, is a cloud port of MinusZoneAI's ComfyUI-Kolors-MZ, released back in August 2024 alongside the rest of BizyAir's Kolors nodes.

Why Kolors, and why it matters that this exists at all

Kolors landed with a genuinely positive initial reception, "Kwai Kolors is really good!" was a common sentiment, and Kuaishou followed up with real supporting tooling like an IP-Adapter implementation, signaling this wasn't a one-off release. But it launched right as the community's attention was shifting toward Flux, and much of the tooling investment that would normally follow a solid model landing well simply went elsewhere instead. The model itself didn't fail on merit, it arrived in a bad window. That's relevant context for anyone finding this node today: Kolors support in ComfyUI exists and works, but it's a smaller, quieter ecosystem than the mainstream one, which is exactly the gap BizyAir's cloud approach is well suited to fill, you don't need a locally maintained node pack and locally downloaded ChatGLM3 weights to try it.

The input, and what it produces

There's exactly one field: text, a multiline string, your prompt. No CFG scale, no token weighting syntax to learn, just plain text handed to ChatGLM3. Output is BIZYAIR_CONDITIONING, BizyAir's typed conditioning socket, which only makes sense wired into a BizyAir_KSampler or BizyAir_KSamplerAdvanced alongside a Kolors BIZYAIR_MODEL loaded elsewhere in the graph, this encoder is architecture-specific and won't do anything useful paired with a non-Kolors model.

Installing it

Through ComfyUI Manager: search "BizyAir", install, restart. Or clone it manually:

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

You need a BizyAir/SiliconFlow API key before this node produces anything, set through the "click to login" prompt on first use. Signup runs through phone-number verification, and it's worth knowing in advance that users outside China have reported friction completing that step.

Common issues

The single most common mistake with this node isn't really an error at all, it's architecture mismatch: wire its BIZYAIR_CONDITIONING output into a sampler running a non-Kolors model and you'll get an unusable or nonsensical result rather than a clean failure, since the conditioning format is specific to how Kolors expects text guidance. Make sure the BIZYAIR_MODEL feeding your sampler is actually a Kolors checkpoint.

As with every node in this pack, a run that stalls or fails outright with no clear workflow-level cause almost always traces back to the API key, check that before anything else. Separately: if BizyAir shows up in your custom_nodes folder without you having installed it, that's a known ComfyUI Manager quirk (misresolving an unrelated missing-node error on an older shared workflow), not malware, and safe to remove through the Manager if you don't want it.