Effort

Effort returns the trace of a model learning from human dance to the body. The training log becomes both music and a sequence of Laban Effort words. The music conditions the trained model and accompanies a human dancer, while the words guide their improvisation as choreographic prompts.

The generated dance and the improvised dance share one training trace and one timeline, but take shape through different forms. As movement passes from bodies into data, through training, and back into bodies, the work asks how agency is distributed across dancers, data, models, and artistic interpretation. Who, or what, is making the dance happen?

When a machine learns to dance from human bodies, what happens between the recorded dancers and the dance it produces? Effort begins with human dance and returns to a human body through a machine. An open-source music-conditioned diffusion model, Lodge [1], is trained on FineDance [2], a motion-capture dataset performed by twenty-seven dancers credited as a group. During training, the changing loss and gradients leave a numerical trace. We take this trace out of the engineering background and turn it into two forms: it becomes music, and it becomes a sequence of Laban Effort words [3] such as Float and Press.

The music conditions the trained model and accompanies a human dancer, while the words become a score for their improvisation, an open choreographic instruction. A generated body, drawn as pale continuous lines from the model’s own joint positions, and a living body then move within the same temporal structure, each meeting the training trace through a different representation. Motion capture reduced the dancers to points in space, and the drawing repeats that reduction.

Human dance performance and improvisation by Qianni Wang.

Existing music-driven dance-generation systems [4, 5] treat training as a means to obtain a model and generated motion. Effort instead treats the numerical trace of training itself as artistic material. Lodge was trained from scratch so that the trace covers the full run, from the first batch to the last. FineDance’s twenty-seven professional dancers were recorded across twenty-two genres, so the trace records that range entering one set of weights.

A change-point detector divides the trace into fourteen phases, and an artist-designed mapping gives one Laban Effort word to each. Laban’s factors describe how a movement is weighted, timed, and directed, and each is paired with the training quantity that plays the same role in optimization. The gradient norm is the force of each update on the weights, read through Weight. Its volatility within a phase shows the evenness of the updates, read through Time. The strength of the temporal trend shows the steadiness of a phase’s direction, read through Space.

A second mapping turns the same trace into a musical composition whose parameters and phase structure can be reproduced from the log. The composition is re-performed with a generative music service [6], which changes its timbre and production while preserving the melody and structure. One cover is heard in the video, and a second cover of the same score conditions the model, so the human dancer and the model share one temporal environment. The model generates a skeletal dance from the music; the dancer hears it, interprets the Effort words, and translates words and sound into movement.

The cycle begins in dance, and the trace returns to dance. Motion capture turned the movement of twenty-seven dancers into numbers, and training turned those numbers into weights. Returning the training trace to a living body runs that path backward, and the vocabulary of that reading, Weight, Time, and Space, comes from dance as well. An Effort word does not prescribe a single movement; it passes through the dancer’s training and embodied judgment before becoming one. Viewers may perceive intention in the generated movement, although the model itself has no expressive intention.

Much like Sol LeWitt’s wall drawings [7], Effort separates the production of an instruction from its bodily realization. Laban used the word “Effort” for the inner impulse behind a movement, and the title gives that word to a training run. A body becomes data. The data becomes a model. The model’s training trace becomes music and words, and the music and words become movement again. After this circuit, whose dance are we watching?