polzounov

Drifting + MeanFlow Hybrid — Full Recipe

A 1-NFE generative model for continuously-conditioned data. One network, two losses.

---

0. One-paragraph summary

Train a MeanFlow backbone u_θ(z, r, t, c) that predicts the average velocity of the flow from time r to time t along the conditional path. The loss has two parts:

polzounov

[
	// Match how literally everything else works (web browsers, preview, files, EVERYTHING!)
	{ "keys": ["super+t"], "command": "new_file" },
	{ "keys": ["super+n"], "command": "new_window" },

	// Add a folder to project (why did they remove this??)
	{ "keys": ["super+shift+o"], "command": "prompt_add_folder" },

	// Find/select all that match current selection (was ctrl+super+g which seems a bit random)
	{ "keys": ["super+shift+d"], "command": "find_all_under" },

polzounov

import numpy as np
from pyrr import matrix44, plane, vector


def rotation_about_x(theta):
    return np.array(
        [
            [1, 0, 0],
            [0, np.cos(theta), np.sin(theta)],
            [0, -np.sin(theta), np.cos(theta)],

polzounov

# for bash completion extention
if [ -f /sw/etc/bash_completion ]; then
   . /sw/etc/bash_completion
fi

# Bash helpers & path stuff
alias reset='source ~/.bashrc'
alias r='source ~/.bashrc'
alias bp='vim ~/.bashrc'

polzounov

import sys


def main():
    # TODO more complicated replacements
    replacements = {
        "->": "$\\rightarrow$",
        "₁": "_1",
        "₂": "_2",
        "ϵ": "\\epsilon",