/series/reading-flow-matching-across-domains/ Recent content in Reading Flow Matching Across Domains on itu Hugo en &copy; 2024 itsublog. All rights reserved. Sat, 10 Oct 2026 02:49:00 +0900 /posts/cs180-pjs/rectified-flow/ Wed, 20 Nov 2024 20:39:29 +0800 /posts/cs180-pjs/rectified-flow/ <p>Here is the (simple) explanation for the framework <strong><a href="https://arxiv.org/abs/2209.03003">Rectified Flow</a></strong>.</p> <p><a href="https://github.com/itsuitsuki/simple-rectified-flow"><strong>A simple implementation</strong></a></p> <h1 id="overview">Overview</h1> <p>Rectified Flow (RF) is a generative modeling method, which tries to transport data from source distribution \(\pi_0\) (which corresponds to the pure Gaussian distribution \(\pi_0=N(0,I)\)) and the target distribution \(\pi_1\), which is the distribution of clean images.</p> <p>The overall objective is to <strong>align the velocity estimate</strong> (using the UNet, denoted as \(v_\theta\) now) <strong>to the actual velocity</strong> between the source image \(X_0\) and the target image \(X_1\). First, the timesteps here are all normalized between \(t\in[0,1]\), instead of spreading in \(\{0,\cdots, T\}\).</p>