Reproduce “pi-Flow: Policy-Based Few-Step Generation via Imitation Distillation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Hansheng Chen, Kai Zhang, Hao Tan, Leonidas Guibas, Gordon Wetzstein, Sai Bi
π-Flow reduces costly network evaluations by having a student predict a rule for continuing denoising, rather than a direct image jump. The rule supplies velocities for integration substeps between network calls. At four NFE, GM-FLUX’s teacher-reference FID is 14.3 versus 23.3 for SenseFlow, while diversity scores are 0.229 versus 0.151; these selected results do not establish superiority on all measures.
Key insights
- Decoupling network evaluations from ODE integration substeps enables both significant inference speedup and the retention of high precision from dense multi-step solvers.
- On-policy imitation learning, as implemented in π-ID, provides a stable and scalable distillation process with robust error bounds, simplifying training compared to prior complex distillation methods.
- GMFlow policies demonstrate superior expressiveness and robustness due to their ability to model dynamic denoising posteriors, efficiently approximating complex ODE trajectories and handling trajectory perturbations effectively.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.