Zinuo You
← All work
open-source2024 – Present

GenODE / GICO

Learning context-conditioned integration clocks for frozen generative ODE and flow-matching backbones.

  • Efficient inference
  • Flow matching
  • Neural ODEs

Problem

At low function-evaluation budgets, a fixed uniform time grid can spend model calls in the wrong parts of a generative trajectory.

Approach

Represent the inference schedule as a continuous density, learn where integration effort is useful, and invert the density into a strictly increasing time grid.

Role

Lead researcher and implementation author

Spend inference where it matters

Generative ODE samplers turn a fixed number of vector-field evaluations into a path from noise to data. Uniform spacing is simple, but at low NFE it treats every part of that path as equally difficult. GICO instead learns a normalized time density conditioned on the solver budget and model context.

Integrating the density gives a monotone cumulative clock. Inverting it produces a strictly increasing grid that concentrates evaluations where the learned policy expects them to be most useful.

Teacher, student, and frozen backbone

The teacher learns schedule utility from evaluated clocks. A student predicts a context-conditioned density that can be applied without changing the frozen generative model. The separation is deliberate: schedule learning improves the inference path while preserving the backbone, solver contract, and target NFE.

Reproducibility is part of the method

The public package treats evaluation and release integrity as first-class research outputs. It includes strict checkpoint validation, locked-test exclusion, canonical reference clocks, deterministic archives, and explicit handling of third-party model and dataset terms.