Video Overlay Removal Flow-Matching Diagram

Overview

A three-panel diagram of a flow-matching video model that removes on-screen overlays such as subtitles. It is an editable PowerPoint diagram adapted from the original paper figure under CC BY 4.0.

What the diagram shows

(a) Flow loop: source video X → VAE encoder E → latent z⁽⁰⁾ = z_s → state z^(k) → velocity predictor (fixed source cues, CFG, instruction c, time τ) → guided v_θ → × Δτ (Δτ < 0) → ⊕ with a current-state bypass → z^(k+1), looped by "update state and time", then VAE decoder D → prediction at step K. (b) Fixed source conditioning: visibility-aware aggregation over frames t − 2 … t + 2 (masked frames crossed out) → visible mean → project → × dilated mask → A; z_s → global mean + MLP → G; mask M → mask embedding → m_tok e_M; summed into the source condition C_s. (c) Shared velocity predictor: z_τ → patchify → ⊕ C_s → N video tokens plus text tokens → a stacked DiT block (joint video–text attention, dense FFN, τ and 3D RoPE, ×24) → v_θ. Every box, arrow, label and math term is a native shape you can recolor, resize and relabel for your own slides, posters or papers.

File format

The download contains .pptx (1 slide, 13.52 × 6.89 in; opens in PowerPoint, Keynote, WPS and LibreOffice Impress; Google Slides works too, but complex freeforms may shift slightly), .svg and .drawio. Includes preview.png (full resolution) and a 1200 px web preview.

Shapes and editing

The file has 812 native editable shapes: 323 auto-shapes, 54 text boxes, 151 freeforms, 284 lines/connectors (2 small groups). No embedded images. Everything is vector. You can edit all text and math (z⁽⁰⁾ = z_s, z^(k), z^(k+1), v_θ, Δτ, z_τ, C_s, A, G, m_tok e_M), the 3D latent grid cubes with masked-region marks, film-strip frames, the encoder / decoder trapezoids, stacked velocity-predictor and DiT cards, token-cube rows, ⊕ / ⊗ nodes, panel headings and every routed connector.

Embedded images and fonts

This file has no embedded images; everything is vector. The source-video and prediction frames inside the film-strip frames are simplified vector placeholder scenes (the source frame carries generic white overlay bars instead of real subtitles); paste your own frames over them if you need real footage. Body text uses Calibri and formulas use Times New Roman. Calibri ships with Microsoft Office and Times New Roman is preinstalled on Windows and macOS (LibreOffice substitutes the metric-compatible Carlito and Liberation Serif). Sub- and superscripts are real formatted text runs, not equation objects. The SVG keeps text as text; the draw.io file is a shape-by-shape conversion and may need small nudges. This is a faithful redraw for reuse, not the authors' original source file.