Video Diffusion KV-Cache Pruning Diagram
Overview
A two-band diagram for KV-cache pruning in autoregressive video diffusion. It is an editable PowerPoint diagram adapted from the original paper figure under CC BY 4.0.
What the diagram shows
The top band shows the online cache update: historical cache H_{<i}, current chunk x_i, current KV, gathering KV with mask m_i and the updated cache for the next chunk. The lower band explains the mask in three steps: compare a probe clean prediction with the final prediction, measure the per-token discrepancy ℓ_{i,p} as a heat map, and threshold it into a retention mask m_{i,p} = 1[ℓ_{i,p} > γ]. 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, 13.85 × 4.39 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 243 native editable shapes: 200 auto-shapes, 32 text boxes, 0 freeforms, 11 lines/connectors. 6 embedded images (see below). You can edit all labels and formulas, every cache cell (history / current / retained / pruned ×), the 1-2-3 step headers, the retention-mask grid, legends and arrows.
Embedded images and fonts
6 items are embedded images: the context frames, the noise sample, the probe and final frames, the discrepancy heat map and its color bar. Text uses Times New Roman, which is preinstalled on Windows and macOS (LibreOffice uses the metric-compatible 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.
Source & license Adapted from Figure 3 of DeCoPrune: Efficient KV-Cache Pruning for Autoregressive Video Diffusion via Denoising Consistency by Zeqi Xiao, Qingle Liu, Kaiwen Zhang, Yifan Zhou, Zihan Ding, Xingang Pan (arXiv:2609.39096, https://arxiv.org/abs/2609.39096). Licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Redrawn as editable shapes; changes made by 叫我Alonzo就好了/alonzolab. Not endorsed by the original authors. Attribution to the original paper is required when you publish or share the figure; the exact credit line is in LICENSE-ATTRIBUTION.txt inside each download.