gabrielelozupone/kokon is a local desktop app that edits photos by describing the desired change in a sentence. It runs from a single installer with no Python installation and no cloud, supports Windows and Mac, and also comes with an MCP server for AI agents based on Qwen-Image 2.1. The November 4, 2026 commit 69fc57e, titled "Video: Kokon 0.12 (Update012, Update012UHD), voice round 3", added 853 lines and deleted 2 lines.
Everything that changed is promotional video production material. A roughly three-minute 0.12 introduction video sequence was newly created, and two video compositing compositions were registered: Update012 and Update012UHD for UHD. The scenes contained exactly the features that 0.12 covers. These include the 레이어스 랩 feature, which separates layers for objects such as letters, bats, ghosts, and spiders, repaints the background, and then hides, moves, or resizes each of them; the feature that expands and completes with AI only the one spider that was cut out; Blend into scene, which reads where the objects in the screen are standing, draws shadows, and adds reflections on wet or glowing floors; the coloring feature where you paint the desired spot yourself and then say what to put there, and Prompt+ converts it into a single detailed request and draws it; and PSD export. The narration was generated sentence by sentence with the same Chatterbox base voice as the 0.10 and 0.11 videos, and 20 sentences of wav files were newly added, including the ending card. Only line 20, the one used for the ending card, got a new slow tone (base, 0.3, 0.2, 0.7) added, inserting a 0.8-second pause between sentences. The comment that the default 0.4 seconds sounded rushed remains as it is.
The video says it runs on its own engine on Mac. The figures written in the script are based on M2 Pro 16GB, 1024², Qwen-Image 2.1 Turbo, 8 steps, and it says that processing a 1-megapixel image takes about 65 seconds (sd.cpp: 276 seconds), four times faster than before. However, this commit changes not app code but only the script, the audio files, and the video timeline, and those figures are the project's own measurements written in its own script. The same script states that Blend into scene is an experimental feature, and that the inpaint result gets pasted as-is on top of the original outside the painted region. In other words, what can be confirmed by this commit reaches only as far as what 0.12 claims; it is not material in which a third party has verified performance and output quality.