Swing Releases Choreo, a Beat-by-Beat Human Motion Generation Model

The Swing team (andrewncarr) has released Swing Choreo, a new human motion generation model, now live on Performance Generator. Users describe a performance beat by beat, and the model executes each motion in sequence while generating hand and facial detail. It supports multi-character interaction in the same scene (e.g., two characters playing rock-paper-scissors) and can move characters to specified positions in world coordinates or camera direction.

Confirmed

• Choreo is live on Performance Generator; for use cases needing fast triage (results in seconds), the previous Swing Sketch remains available
• With no decent public benchmark in human motion generation, the team built their own, evaluating motion sequences, hands, face, multi-character coordination, directionality, and physical plausibility; Choreo scored 61/100 while the runner-up (an open-source research model, Sy) scored just 38
• In blind tests with professional animators (with randomized clip order), 88% overall preferred Choreo over Swing Sketch; by category: 100% for multi-character interaction, 94% for hands and fingers, 92% for multi-motion sequences, and 86% for spatial direction
• Comparison tables show Choreo leading across multi-motion sequences, finger motion, facial expressions, multi-character interaction, world/camera direction control, long-prompt adherence, and motion diversity — but at the cost of speed, with each generation taking several minutes
• The team's "motion quality vs. per-generation time" curve shows Choreo at the frontier across 25%, 50%, and 100% compute budgets, surpassing commercial models and open-source models A/B; the author stresses most gains come from data selection and post-training rather than brute-force compute
• Training followed an iterative loop: capture motion data → train → animator evaluation → feed model errors back into the next round of data collection, iterating continuously over the past year, with each round improving quality and guiding the next data collection step

Why it matters

• Joint generation of fingers, face, and body motion plus multi-character interaction has been a long-standing pain point in animation production that couldn't be solved end-to-end; Choreo makes it possible to produce an entire choreography from a single description
• The animator-review data feedback loop and the quality-vs-compute frontier driven by data selection and post-training offer a reference engineering methodology for motion generation; the purpose-built benchmark with public dimensions also partially fills an evaluation gap in the field

2026-10-13 ~ 2026-10-14 · 10 related posts

• Swing Choreo generates multi-person human motion from beat-by-beat descriptions — andrew_n_carr · 2026-10-13
• [source] Swing launches Choreo, a text-to-motion model that performs whole sequences from one prompt — andrew_n_carr · 2026-10-13
• Swing Choreo generates finger articulation and facial motion alongside body movement — andrew_n_carr · 2026-10-13
• [source] Animators prefer Swing Choreo 88% of the time in blind tests, 100% on multi-person scenes — andrew_n_carr · 2026-10-13
• Swing Choreo beats open and commercial text-to-motion tools on every capability except speed — andrew_n_carr · 2026-10-13
• Swing shows compute-quality frontier for motion models: gains came from data selection, not scale — andrew_n_carr · 2026-10-13
• Swing details Choreo's training loop: animator evals feed back into motion-capture data — andrew_n_carr · 2026-10-13
• [source] No public benchmark for motion generation, so Swing built one: Choreo scores 61 vs 38 — andrew_n_carr · 2026-10-13
• Swing Choreo launches with a self-built motion benchmark: 61/100 vs 38 for the next best — andrew_n_carr · 2026-10-13

1 near-duplicate retellings: andrew_n_carr