Generative Vision Atlas

Research line · ascendant

Inversion for flow models

Recover the noise that produced a real image so it can be re-generated with a change, and control the error that recovery introduces.

What defines membership

Editing a real image requires first mapping it back into the generator's own trajectory, and doing that accurately is a numerical problem in its own right.

How the line developed

Read top to bottom: what came before, the idea itself, the evidence for it, what improved, and where it breaks.

The idea

RF-Inversion · 2024-10core

Treats inversion as optimal control, equivalent to a rectified stochastic differential equation, to steer against drift.

Improvement

RF-Solver · 2024-11core

Isolates per-step truncation error and replaces Euler integration with a high-order expansion.

FireFlow · 2024-12strong-followup

Targets the cost of precision, claiming second-order accuracy at first-order expense.

What it gets right

  • Necessary for editing real photographs rather than generated ones
  • Straight-path flow objectives make the problem more tractable than it was for diffusion

Where it is weak

  • Straighter paths reduce error without eliminating it
  • Three papers within three months, none treating the problem as settled
  • Inversion is lossy, and the loss shows up as over-smoothing