Multi-ControlNet Curve Coordinator

Most multi-ControlNet workflows have a subtle problem: you stack a pose net and a depth net, give both a start and end percent, and hope they don't trip over each other. This node attacks that directly - up to four ControlNet slots, each with its own timing window, its own strength curve, and its own preset, all generated in one place and visualized on a single combined graph so you can see the handoffs. It's the pack's answer to "when does each ControlNet get to talk?"

How it works

Each slot (A through D) is a complete mini-scheduler: enable_slot_X turns it on, start_percent_X / end_percent_X define when in the generation that ControlNet is active, preset_X or the manual trio (start_strength_X, end_strength_X, curve_type_X, curve_param_X) define how strong it is across that window. num_keyframes applies to all slots at once so the four streams line up. The node builds a TimestepKeyframeGroup per enabled slot and returns them as four separate outputs - SLOT_A through SLOT_D - each wired into its own Apply Advanced ControlNet instance.

The feature that earns the name "coordinator" is combined_graph: one chart plotting all four strength curves, with the active windows shaded, so a "smooth handoff" is something you can verify at a glance instead of trusting. The info output is the text summary.

The shapes of a good schedule

The README's use cases are worth internalizing because they're the actual reasons this node exists:

• Sequential control - pose guides 0–40%, then depth takes over 40–100%. No overlap, no argument.
• Overlapping transitions - give the two curves overlapping windows and opposite directions (pose fading out while depth fades in) for a smooth handoff instead of a cut.
• Targeted bursts - a slot enabled only for a 20% window where a specific structure problem lives.

Installing it

Same pack, same Advanced ControlNet dependency:

cd ComfyUI/custom_nodes
git clone https://github.com/diffussy69/comfyui-curved_weight_schedule
pip install matplotlib pillow numpy torch scipy

The outputs are TimestepKeyframes, so Kosinkadink's ComfyUI-Advanced-ControlNet must be installed or you'll get the familiar import error. Restart, hard-refresh.

Where people get burned

The divider inputs (divider_a etc.) are just visual separators - they do nothing, which is easy to forget when a workflow insists on them. Bigger trap: four slots all with default 0–1 windows is four ControlNets fighting for the whole generation, which is how you get mush. The whole point is giving each slot a different window. Start with two slots and hand-check the combined graph before adding more - the graph is the feature; use it.