SCAIL Beat-Driven Dance

This is the heart of the pack - the node the README calls "the Choreographer." SCAILBeatDrivenPose takes a base skeleton, a beat structure and audio features, and generates a full physics-based dance sequence with 80+ poses across styles like hip hop, rock and disco. Where the AIST/CMU nodes replay real motion capture, this one synthesizes dance in real time, which means it's fully audio-reactive: bass drives the bounce, treble drives the arms, and moves snap onto the beat with adjustable anticipation. It's also the only node in the pack that does real multi-character coordination out of the box.

How it works

The movement is driven by a custom physics engine rather than a lookup table. A few design choices make it worth understanding:

• Sticky feet. Ankles get zero drag and double stiffness in the engine's MotionDynamics model, forcing them to snap to the floor unless a specific move lifts them. That's the difference between a grounded dancer and a floating puppet - the classic failure of procedural animation.
• Neutral structure init (auto-rigging). If your reference pose has the character mid-stride or on one leg, the engine computes a neutral standing skeleton from their bone lengths and interpolates Reference → Neutral → Dance move. Without this, a character whose reference image has one foot up stays stuck with one foot up all dance.
• Scale-aware movement. Amplitudes are normalized to torso length, so a background character moves with appropriately smaller absolute motion and won't collide with the foreground dancer.

The inputs are many but they divide into two buckets. Structure: base_pose, beat_info, audio_features (all wired in - from SCAILBasePoseGenerator or SCAILPoseFromDWPose, SCAILBeatDetector, and SCAILAudioFeatureExtractor). Feel: the knobs that actually change the output:

• dance_style (22 options, auto picks by energy) and energy_style (auto/low/medium/high)
• interaction_mode - unison (everyone identical), mirror (left/right swapped around the leader), random (independent moves). This is where 2–5 characters get interesting.
• motion_smoothness (lower = snappier, higher = floatier), anticipation (frames the move starts before the beat), groove_amount (continuous hip sway), bass_intensity (vertical bounce), treble_intensity (arm/hand jitter)
• phrase_bars (4/8/16 - when choreography changes), staging_mode (off/subtle/dynamic), hit_easing (linear/ease_out/bounce/elastic/punch), pose_blend, loop_mode (force a seamless loop), seed

Output

One output, pose_sequence, into SCAILPoseRenderer, then the rendered frames feed your Wan pose-control nodes. The pack's example workflows run exactly this chain with the SCAIL-preview Wan model.

Installing and common gotchas

Standard pack install: clone ckinpdx/ComfyUI-SCAIL-AudioReactive into ComfyUI/custom_nodes, pip install -r requirements.txt, restart (ComfyUI Manager works too). No downloads for this node itself.

The gotchas are the frame-count mismatch (beat_info and audio_features must come from matching frame_count/fps), and the initial-pose trap - if your reference image has the character in a weird stance, the neutral-structure init handles it, but you'll still get better results starting from active_idle than from a mid-stride frame. And if the dance feels robotic, lower motion_smoothness before you blame the physics.