☁️BizyAir Controlnet Union SDXL 1.0

xinsir's SDXL ControlNet Union is, by a real margin, the most capable single ControlNet artifact in the ecosystem - one Apache 2.0 checkpoint that replaced what used to be a dozen separate ~1.5GB models, released back in July 2024 and still the SDXL ceiling nobody's matched since. This node is BizyAir's all-in-one wrapper around it: instead of loading a checkpoint, a union ControlNet, a type selector, and a KSampler separately, you get one node that takes a prompt and up to six condition images and hands back a finished image. No local weights, no VRAM math - it runs on BizyAir's cloud and bills your account per call.

How it's different from a normal ControlNet workflow

This isn't the modular ComfyUI pattern of load-checkpoint → apply-controlnet → sample. It's a packaged pipeline node - the SDXL checkpoint, the union ControlNet, and the sampler are all baked in behind the scenes. What you get exposed is seed, num_inference_steps (1–50, default 20), num_images_per_prompt (1–4), and guidance_scale (0–100, default 5 - note this sits well below the classic SD 1.5-era "CFG 7" instinct, which tracks what the 2025-era union models actually publish as their recommended range). What you don't get is a choice of base checkpoint, LoRA slots, or a custom sampler/scheduler - you're trading fine control for a one-node setup.

The inputs that matter

Six optional image slots, each one a condition family the union model understands: openpose_image, depth_image, hed_pidi_scribble_ted_image (soft-edge family), canny_lineart_anime_lineart_mlsd_image (hard-edge family), normal_image, and segment_image. You don't have to fill all six - wire in whichever condition images you have, leave the rest empty. prompt (default is literally "a car", the model's own demo string) and negative_prompt (default "watermark, text") round out the text side. control_guidance_start/control_guidance_end (0–1 each) let you limit the ControlNet's influence to a portion of the denoising steps, same as the standard start/end pattern on any ControlNet node.

Output is a single IMAGE - no intermediate latent, no conditioning to inspect. What goes in is what comes out.

Installing it

ComfyUI Manager: search BizyAir, install, restart. Manual:

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

Restart after. No models to download - this node doesn't touch your disk at all, which is arguably the whole pitch of running xinsir's union through BizyAir instead of loading the real checkpoint locally. You do need a BizyAir account and API key configured before this or any other node in the pack will run.

Where people get burned

Expecting per-condition strength control. The real xinsir union checkpoint supports per-condition scaling and stacking multiple conditions with individual weights; this pipeline node collapses that down to one guidance_scale and one start/end window for everything you've plugged in. If you need that granularity, you want the modular BizyAir controlnet nodes (loader + BizyAir SetUnionControlNetType), not this pipeline.

Not knowing which image goes where. The six inputs group multiple related preprocessor outputs under one slot - canny, lineart, anime lineart, and MLSD all share canny_lineart_anime_lineart_mlsd_image, for instance. Feed it the matching preprocessor output, not a raw photo.

Treating guidance_scale like classic CFG. Starting it at the old SD 1.5 default of 7 tends to overcook a union model tuned for a lower default - the node's own default of 5 is closer to what's actually recommended.