☁️BizyAir CogView4_6B_Pipe

Skip the checkpoint loader, skip the CLIP loader, skip the conditioning nodes - this one node is a whole text-to-image pipeline by itself. Type a prompt, get back an IMAGE. No BIZYAIR_MODEL wire to manage, no separate encode step; it's a "pipe" node in the sense ComfyUI regulars use the word: everything a normal multi-node workflow would wire together, collapsed into one box that calls out to BizyAir's cloud and hands you a finished picture. BizyAir builds a handful of these dedicated wrapper nodes for specific popular releases as they land - this one's for CogView4-6B, added to the pack per SiliconFlow's own changelog not long after the model shipped.

CogView4 is Zhipu's (now branded Z.ai) open text-to-image model - the same lab whose visual lineage runs back through CogView and CogVideoX, one of the longer-running "ship it open, MIT or Apache 2.0, no revenue cap" postures in this space. If you're used to Flux or SDXL, CogView4 is a peer in that category: a real, independently-trained base model rather than a finetune of something else, and it's earned enough of a reputation to get name-checked in comparison posts against Flux itself.

Inputs that matter. prompt is a multiline string - natural language, the node's own default example is a full descriptive paragraph about a sports car, so lean toward Flux-style prose rather than comma-separated tags. width and height default to 1536×832 (a widescreen-ish frame) and go up to 4096 on either side in steps of 16; num_inference_steps defaults to 30 (range 1–100). seed is the usual reproducibility knob. num_images_per_prompt (1–4) batches multiple outputs from one call instead of running the node repeatedly.

guidance_scale (default 3.5, range 0.1–100) is the one to actually think about. That default - and the field's name - line up with how Flux exposes its own embedded guidance rather than with classic two-pass CFG on an SD 1.5/SDXL model, where the sane range is 5–9 and the default is 7. Treat it accordingly: nudging it up tightens prompt adherence but flattens the image the way Flux's guidance does at the high end, rather than behaving like SDXL's CFG dial. Start near the default and adjust in small steps rather than jumping to 7 out of SD-era habit.

Output: a plain IMAGE - no BizyAir-specific type here, so it drops straight into whatever comes next in your graph: Save Image, an upscaler, a post-process chain, anything.

Install: search "BizyAir" in ComfyUI Manager and click install, or cd ComfyUI/custom_nodes && git clone https://github.com/siliconflow/BizyAir.git then restart, or comfy node install bizyair via the Comfy CLI - one pack install covers this node along with everything else BizyAir ships. Then register a BizyAir account and set your API key on first use; without it this node, like every other one in the pack, simply won't run the generation. There's no separate model download for CogView4 itself - you're not fetching a multi-gigabyte checkpoint, the 6B model runs on SiliconFlow's infrastructure.

Troubleshooting: if the node fails outright or hangs, check your API key first - that's the most common cause of a stalled or silently-failing BizyAir call, and it applies here the same as anywhere else in the pack. If your images look flat or oversaturated, treat guidance_scale the way you'd treat Flux guidance rather than SDXL CFG - pull it down before you start rewriting the prompt. And because this is an all-in-one node with no exposed sampler or scheduler, there's nothing to tune beyond what's on the widget list here; if you need finer control than that, you'd be looking at a manual checkpoint+CLIP+conditioning chain instead, assuming BizyAir exposes CogView4 that way in the rest of the pack.