The recorded flow data and visualization are from Chongyi Zheng, Zhiyuan Zhou, Qiyang Li, Benjamin Eysenbach, and Sergey Levine’s Value Forcing project: https://chongyi-zheng.github.io/value-forcing/

Licensed under Creative Commons Attribution-ShareAlike 4.0: https://creativecommons.org/licenses/by-sa/4.0/

The video adapts the website’s recorded denoising evolution into a single-row layout with no interactive controls. For thumbnail readability, it shows 14 evenly spaced recorded paths (including the outermost paths) per method, with thicker strokes, moving endpoints, darker colors, and the group labels “Value Forcing” and “Previous Methods.” The four-second loop plays the original evolution at twice its original speed. The original distributions are retained; axis ticks and grid lines are omitted. The adapted video and rendering source are shared under the same license. Run render_teaser.py with Python, Pillow, and ffmpeg to regenerate denoising.mp4 and denoising-poster.png.