The authors systematically compared x1-, v-, and x0-prediction in flow matching, training all against the same denoising target with uniform weighting to isolate parameterization effects.
Spectral hybrid parameterizations robustly accelerate flow matching optimization
The authors identify signal-to-noise ratio and architecture-induced compression as key factors, and introduce parameterizations that adapt across time and data covariance directions.
Research Lab
Ségolène Martin · Anne Gagneux · Quentin Bertrand · Rémi Emonet · Mathurin Massias
Inria · ENS de Lyon · CNRS · Université Claude Bernard Lyon 1 · Université Jean Monnet Saint-Étienne
Research Digest··2 min read
The paper studies why different prediction targets (x1, v, x0) in flow matching yield different performances, and identifies two drivers: signal-to-noise ratio per data covariance direction and neural architecture bottleneck.
Why this paper
From Inria and 6 others
In one line
Hybrid spectral parameterizations for flow matching that adapt across time and data covariance directions accelerate optimization with no extra training cost.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors (3 noted)
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§