A Time-Reparameterized Cumulative Intensity Extrapolation Sampler for Discrete Flow Matching

2026-06-23 · Source: Takara TLDR - Daily AI Papers · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, medium

Summary

Feiyang Fu and Hehe Fan propose the Time-Reparameterized Cumulative Intensity Extrapolation (TR-CIE) sampler to enhance sampling quality in Discrete Flow Matching (DFM) models, particularly when function evaluations (NFE) are restricted. DFM provides a principled framework for generative modeling on discrete state spaces using continuous-time Markov chain dynamics, but efficient sampling under limited NFE has been a challenge. TR-CIE addresses this with two components: a schedule-based time reparameterization that rescales the time grid to mitigate stiffness, and a cumulative-intensity extrapolation updating rule that reuses cached model outputs to improve stepwise intensity approximations. The resulting sampler maintains efficiency, requiring one NFE per step without additional model evaluations compared to standard $\tau$-leaping. Extensive experiments on synthetic tasks, text generation, and text-to-image benchmarks demonstrate improved sampling quality under limited NFE.

Key takeaway

For Machine Learning Engineers developing discrete generative models, if you are constrained by function evaluations, consider implementing the TR-CIE sampler. This method improves sampling quality for Discrete Flow Matching models without increasing computational cost, requiring only one NFE per step. You can achieve better results in text generation and text-to-image tasks, making your models more efficient for deployment.

Key insights

TR-CIE improves discrete flow matching sampling quality under limited function evaluations by reparameterizing time and extrapolating cumulative intensities.

Principles

• Time reparameterization mitigates stiffness.
• Reusing cached outputs improves approximations.
• One NFE per step maintains efficiency.

Method

TR-CIE rescales the time grid based on the noise schedule and reuses cached model outputs from the previous step to extrapolate cumulative intensities on a non-uniform time grid.

In practice

• Apply TR-CIE for efficient DFM sampling.
• Improve text generation with limited NFE.
• Enhance text-to-image benchmarks.

Topics

• Discrete Flow Matching
• Generative Models
• Sampling Efficiency
• Time Reparameterization
• Text Generation
• Text-to-Image

Code references

• gstoica27/DeltaFM

Best for: Research Scientist, NLP Engineer, Computer Vision Engineer, AI Scientist, Machine Learning Engineer

Related on AIssential

• One-step Language Modeling via Continuous Denoising — Continuous denoising via flow-based models can surpass discrete diffusion in language model quality and speed.
• TFTF: Training-Free Targeted Flow for Conditional Sampling — TFTF enables training-free conditional sampling in flow matching by combining importance sampling with stochastic flow and SMC resampling.
• Branching Flows: Discrete, Continuous, and Manifold Flow Matching with Splits and Deletions — Branching Flows enables generative models to naturally handle variable-length sequences by stochastically evolving elements over binary trees.
• Temporal Difference Learning for Diffusion Models — Temporal difference learning improves diffusion model consistency and sample quality, especially for few-step sampling.
• Flow Matching with Semidiscrete Couplings — Flow matching trains generative models by aligning velocity fields with target data through ODE integration.
• DRIFT: A Residual Flow Adapter for Decoding Continuous Outputs in Vision-Language Models — DRIFT adapts VLMs for precise continuous outputs by refining coarse predictions with a flow matching-based residual generative module.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Takara TLDR - Daily AI Papers.