☁️BizyAir Apply ControlNet

Despite the plain display name - just "Apply ControlNet," no qualifier - this is BizyAir's current, recommended ControlNet node. The sibling with "(OLD)" in its name is the one to avoid for new work; this is the one that actually matches how ControlNet gets used in 2026. It applies your ControlNet's spatial guidance to positive and negative conditioning separately, and it gives you a window - a start and end percentage of the sampling process - for exactly when that guidance is active.

Why the positive/negative split and the step window both matter. ControlNet conditions generation on structure pulled from a reference image (edges, depth, pose, whatever the loaded model was trained on) while your prompt still decides content - "the prompt handles what, the ControlNet handles where." Applying it to positive and negative separately, rather than to a single merged conditioning, is just how the modern apply pattern works, and it's what lets this node's start/end percent do something meaningful for both. The step window itself is the more consequential control: the community's standing advice, dating back to ControlNet's earliest release and still true on every current architecture, is to release the ControlNet once composition has formed rather than holding it through the entire sample - start at 0, end around 0.5, and the model gets to add its own detail in the back half instead of being spatially pinned the whole way through. The modern union ControlNets (the single-checkpoint models covering canny, depth, pose and friends on current base models) also publish lower recommended weights than the SD-era default of 1.0 - think 0.65–0.9 rather than maxing this out - so the strength default here is a reasonable starting point, not a floor.

Inputs that matter. positive and negative, both BIZYAIR_CONDITIONING, required. control_net (BIZYAIR_CONTROL_NET, from a BizyAir ControlNet loader) and image, both required. strength (default 1, range 0–10, step 0.01) is the overall weight; start_percent and end_percent (both 0–1, step 0.001, defaulting to 0 and 1 - meaning "on for the whole sample" out of the box) are the step window. Move end_percent down before anything else if your outputs feel over-constrained or your ControlNet seems to be fighting the prompt for fine detail.

Outputs: positive and negative, both BIZYAIR_CONDITIONING - this node replaces your existing positive/negative pair with ControlNet-augmented versions, so wire both outputs into your sampler/guider the same way you would the originals.

Install: search "BizyAir" in ComfyUI Manager and install, cd ComfyUI/custom_nodes && git clone https://github.com/siliconflow/BizyAir.git then restart, or comfy node install bizyair on the Comfy CLI - one pack install brings in this node along with every ControlNet loader and apply variant BizyAir ships. It's a cloud call like the rest of the pack, so a registered BizyAir account with your API key set is required before it'll run - the README flags this as a first-use step, and skipping it is the most common reason a BizyAir node silently fails.

Troubleshooting. If the ControlNet's influence feels too strong or too literal - flattening detail, fighting the prompt - pull end_percent down toward 0.4–0.6 before you touch strength; letting the condition run the whole sample is a much more common cause of over-constrained output than the weight being too high. If it feels too weak, check that your reference image actually matches what the loaded control_net expects as a preprocessed condition (a depth ControlNet fed a raw photo instead of an actual depth map won't do much). And as with any BizyAir node, a stall or generic failure notice is worth an API-key check before you go debugging the graph itself.