CuCaT: Cut-Structured Camera Trajectory Generation for Dance Music Videos

Overview of CuCaT: music and dance inputs, cut and transition selection, cut-aware shot plan generation, and the resulting camera trajectory.

Cuts renew the view.
Shots carry the motion.

A dance camera has to keep the dancer visible, renew the view at cuts, and stay smooth within each shot. CuCaT models cuts and within-shot motion as separate events, each described by a performer-relative shot plan that artists can read and edit.

On the DCM benchmark it gives the best continuity and variation, keeps the dancer in frame almost always, and comes close to artist-made cameras in a user study.

Supplementary video · 4:39 · with sound

01 · The observation

Three properties of a dance camera

A rendered dance camera holds up only if three things are true at once. The key observation is that they live on different temporal scales, so a single homogeneous output stream can't satisfy all of them.

per frame

Visibility

The dancer stays in frame across the whole performance. This is a constraint defined relative to the performer's body in every frame.

discrete event

Variation

New angles and framings keep the view from turning into one monotonic orbit. Variation happens at cuts, where the viewer's relation to the dancer resets.

segment level

Continuity

Motion inside a shot stays smooth enough that viewers can follow the choreography without fighting the camera. Continuity is defined only between cuts.

◆ CutAbrupt viewpoint change

● TransitionSmooth motion within a shot

Both, in one trackCuCaT output

02 · Representation

Semantic shot plan

Every event is a categorical shot plan defined relative to the performer. A deterministic solver turns it into an MMD camera. Try it below: each frame is a real render.

SABA render of the Okayu model for the selected shot plan

real SABA render · fixed pose

03 · Pipeline

Method overview

Cut-structured keyframe selection

Pick cuts as frame pairs (f−1, f), then transitions inside each shot.

CNN + Transformer

Role-aware shot-plan generation

Separate cut and transition heads on a shared Transformer.

separate categorical heads

Trajectory optimization

CP-SAT picks one candidate per event, avoiding bounces and weak cuts.

CP-SAT, discrete

Interpolation & MMD solve

Predict easing per keyframe and solve each plan into an MMD camera.

MLP + solver

04 · In action

Example output

Predicted cuts and transitions for a DCM test clip, with the shot plan in effect.

CuCaT output · DCM test clip C_8_1

cut

cuttransition· music waveform

05 · Results on DCM

Quantitative results

DCM benchmark, against dense, keyframe and token-based baselines.

Continuity · FID_k

1.655

−48.9% vs. best baseline

Variation · FID_s

0.013

~20× lower

Visibility · DMR

0.03×10⁻³

~67× lower

User study · win rate

80.5%

vs. DanceCamAnimator

	Continuity		Variation		Visibility
Method	FID_k ↓	Dist_k ↑	FID_s ↓	Dist_s ↑	DMR ↓	LCD ↓
Ground truth	–	3.275	–	1.731	1.42×10⁻³	–
DanceCamera3D	3.749	1.631	0.280	1.326	2.50×10⁻³	0.147
DanceCamAnimator	3.453	3.140	0.268	1.293	2.20×10⁻³	0.152
TemMEGA^‡	3.237	1.961	0.255	1.347	2.00×10⁻³	0.141
CuCaT (ours)	1.655	3.172	0.013	1.620	0.03×10⁻³	0.184

Dist: closer to GT is better. ‡ Numbers from the original paper.

User study · N = 21, 1–7 scale

CuCaT GT (artist) DanceCamera3D DanceCamAnimator

1234567

Forced choice vs. CuCaT

Ablation study

each row removes or replaces one component

Variant	FID_k ↓	Dist_k ↑	FID_s ↓	Dist_s ↑	DMR ↓	LCD ↓
Full CuCaT	1.655	3.172	0.013	1.620	0.03×10⁻³	0.184
Representation
Unified event	2.011	2.593	0.046	1.336	1.90×10⁻³	0.211
Numerical camera rep.	6.025	1.243	0.788	1.427	66.0×10⁻³	0.261
Training
w/o budget allocation	3.104	3.664	0.012	1.621	0.40×10⁻³	0.181
No cut-conditioning	5.257	1.687	0.0059	1.663	0	0.186
Shared heads	1.889	2.660	0.0040	1.662	0.30×10⁻³	0.192
Losses
w/o pre-cut contrast	5.017	1.730	0.012	1.644	0	0.183
w/o smoothness	1.798	3.284	0.011	1.637	0.03×10⁻³	0.182
w/o distance–FoV joint	4.250	1.872	0.343	1.659	0.33×10⁻³	0.228
w/o visual-scale	1.662	2.905	0.016	1.624	0	0.184

Unified eventCuts blur into keyframes (FID_s 3.5×)

Numerical rep.Worst overall, dancer often lost

No cut-conditioningContinuity drops (FID_k 5.26)

06 · Qualitative

Qualitative results

Baseline on the left, CuCaT on the right. Same dance, same music.

Compare with

Song

DanceCamAnimatorCuCaT (ours)

07 · Citation

Cite CuCaT

@inproceedings{lim2026cucat,
  title     = {CuCaT: Cut-Structured Camera Trajectory Generation for Dance Music Videos},
  author    = {Lim, SooBin and Noh, SeongRae and Kwon, Jeongyoun and Kim, YoungBeom and Kang, HyeongYeop},
  booktitle = {SIGGRAPH Asia 2026 Conference Papers (SA Conference Papers '26)},
  year      = {2026},
  publisher = {ACM},
  address   = {Kuala Lumpur, Malaysia},
  doi       = {10.1145/3829340.3842347}
}