Advanced Curved ControlNet Scheduler

ControlNet's single most useful parameter is when it's active - start it at 0 and end it at 0.5 and the condition locks in composition early, then lets the model improvise detail. Most of the ecosystem gives you two fields for that. This node gives you a curve: ControlNet strength that fades, spikes, oscillates, or bell-curves across the whole generation, defined by presets, easing functions, custom math formulas, or even a mouse-drawn shape. It's the flagship of the curved_weight_schedule pack, and the one node the author clearly spent the most time on.

How it works

Underneath, it's a TimestepKeyframe generator for Kosinkadink's ComfyUI-Advanced-ControlNet. The node samples a strength curve across num_keyframes points between start_percent and end_percent, then emits a TIMESTEP_KEYFRAME object that Apply Advanced ControlNet consumes - each keyframe holding the curve's strength at that point in the schedule. Pick a preset (Fade Out, Fade In, Peak Control, Valley Control, Oscillating, and five more) and it fills in the strength and curve shape for you. Pick Custom and you're driving it manually.

The genuinely useful knobs:

• preset - nine one-click shapes. If the UI isn't updating fields when you select one, the values still apply internally; the JS extension just makes it visible.
• mode / total_steps - percent (0–1) or steps (absolute sampler steps).
• num_keyframes, start_percent, end_percent - the skeleton of the schedule.
• start_strength, end_strength, curve_type, curve_param - the manual curve. curve_param controls steepness; higher is more extreme.
• custom_formula - math with t from 0 to 1: t**2, sin(t*3.14), 1-exp(-5*t). Whitelisted functions only - no imports, no exec.
• batch_masks - the temporal-masking hook. Wire a mask batch here and keyframe i carries mask i.
• blend_curve_type / blend_amount, mirror_curve, repeat_curve, adaptive_keyframes - the curve-modulation toys. Fun, optional.
• comparison_curve - overlay a second curve on the graph for A/B testing. save_curve / curve_filename dump the curve to CSV.

Outputs: TIMESTEP_KF (into Apply Advanced ControlNet's timestep_kf), curve_graph (an IMAGE - preview it and you'll see your schedule), and curve_stats (average, min, max, area under the curve as text).

Installing it - read this twice

This pack sits on top of Advanced ControlNet, and the temporal features need the modified fork, not the original:

cd ComfyUI/custom_nodes
git clone https://github.com/diffussy69/comfyui-curved_weight_schedule
git clone https://github.com/diffussy69/ComfyUI-Advanced-ControlNet
cd ComfyUI-Advanced-ControlNet && git checkout per-keyframe-images && cd ..
pip install matplotlib pillow numpy torch scipy

Restart, then hard-refresh (Ctrl+Shift+F5). Skip the fork and non-temporal scheduling still works - but batch_masks silently goes nowhere, which is the pack's headline feature.

Where people get burned

The Could not import TimestepKeyframe error means Advanced ControlNet isn't installed (or wasn't restarted). Keyframes "all starting at 0%" is expected behavior from many schedulers - chain the pack's Redistribute Keyframe Percents node after this to spread them. And honest take: this pack has essentially zero community footprint - nobody's sharing battle-tested workflows, so you're largely on your own tuning it. The graph preview and stats outputs are your allies; check them before you trust a curve. Start with the plain Curved ControlNet Scheduler first if you just want a simple fade - the Advanced version is a toolbox, and you'll only need half of it.