Edit photos by describing the change, on your own GPU. A free, open-source alternative to Nano Banana and ChatGPT image editing: photos stay on your computer, no daily limit. One installer for Windows, Mac and Linux sets up the best open model (Qwen-Image 2.1). MCP server for AI agents.
TypeScript ai-photo-editorcomfyui-alternativegenerative-fillimage-editingimage-generationinpaintinglinuxlocal-aimacosmcpmcp-servernano-banana-alternativeofflinephoto-editorprivacyqwen-imageqwen-image-2-1stable-diffusion-cppvulkanwindows
0.10.2 released: STATUS, README roadmap, site version
Gabriele Lozuponee411d9672dba
diff --git a/README.md b/README.mdindex 1fd882d..3db6281 100644--- a/README.md+++ b/README.md@@ -700,6 +700,7 @@ or macOS version and the log from `KokonData/logs` (see the table above).- [x] Qwen's official 8-step Turbo model for every preset- [x] Prompt+: the prompt enhancer on the model Kokon already has (0.10.0)- [x] Your own LoRA files, per project (0.10.0, beta)+- [x] New images from reference photos (a person, a product, a style), with Prompt+ (0.10.2)- [x] AMD GPUs through Vulkan (beta; RX 7900 XTX and RX 9060 XT reported working, RX 6000 cards still slow)- [ ] **0.11:** the interface in the browser, served by Kokon itself, on your local network with a password;then a Docker image for NVIDIA ([#15](https://github.com/GabrieleLozupone/kokon/issues/15))diff --git a/docs/STATUS.md b/docs/STATUS.mdindex 7d0b39d..13c7aff 100644--- a/docs/STATUS.md+++ b/docs/STATUS.md@@ -3,7 +3,7 @@**Updated 2026-10-23.** What is in progress, what comes next, and where help is welcome. Kept current at the endof every task (see [AGENTS.md](../AGENTS.md#keeping-docsstatusmd-current)); GitHub is the source of truth.-## Current version: [0.10.1](https://github.com/GabrieleLozupone/kokon/releases/tag/v0.10.1) (2026-10-22)+## Current version: [0.10.2](https://github.com/GabrieleLozupone/kokon/releases/tag/v0.10.2) (2026-10-23)Windows, macOS (Apple Silicon, 15+) and Linux (`.deb`, `.rpm`). Installed copies update from the app.@@ -16,20 +16,6 @@ Tracked in [#15](https://github.com/GabrieleLozupone/kokon/issues/15); design in- A Docker image for NVIDIA (CUDA) first; an AMD/Intel (Vulkan) image once it can be tested.- Later: the desktop app as a client of a Kokon server elsewhere.-## Being tested: 0.10.2 (on master)--Notes: [release-notes-0.10.2.md](../marketing/release-notes-0.10.2.md). Tested on Windows (CUDA); macOS and-Linux next. The new engine already reaches 0.10.x apps through `engine.json`.-- **New image from reference images**: with New image, the images you add are references (a person, a- product, a style) for a new picture sized by your quality and aspect ratio, not by image 1; Prompt+ has its- own instructions for it. Also in MCP (`generate_image` with `reference_images`) and the HTTP API.-- **Prompt+ menu**: off, write and make, or write the prompt into the box first to read and change it.-- **Longer Prompt+ prompts** for crowded scenes (they were cut at 512 tokens).-- **Edits never reuse their source's seed** (that made them over-sharpened and crunchy).-- **Vulkan decodes in 512 px tiles**, under AMD's 2 GiB buffer limit (decodes took 25-70 s on AMD cards).-- **New engine** [kokon-1013-1814814](https://github.com/GabrieleLozupone/stable-diffusion.cpp/releases/tag/kokon-1013-1814814)- (upstream fd21ecf, with the Vulkan fix for cards without Resizable BAR).-## Next- **0.12: reliability foundations** (details soon).@@ -42,8 +28,6 @@ Linux next. The new engine already reaches 0.10.x apps through `engine.json`.- **AMD RX 6000 (RDNA2) cards are slow on Vulkan** (an RX 6700 XT takes 21-41 s per step at 1 MP). Underinvestigation: flash attention and memory placement are ruled out (the model stays in VRAM, the GPU is at 99 %).RX 7900 XTX and RX 9060 XT work well.-- **Slow VAE decode on AMD cards on Vulkan**: the decode asks for a buffer over the driver's 2 GiB limit and takes- 25-70 s instead of a few seconds. Fixed in 0.10.2.- **LoRA is beta** and tested on Windows only; reports from macOS and Linux are welcome([#33](https://github.com/GabrieleLozupone/kokon/issues/33) has a product-shot test).- **Macs are slow** compared with NVIDIA cards (about 4 minutes for 1 MP on an M2 Pro 16 GB). Engine speed-ups are@@ -68,10 +52,12 @@ Linux next. The new engine already reaches 0.10.x apps through `engine.json`.## Recently shipped-- **On master** (next release): `kokon-mcp --help` ([#36](https://github.com/GabrieleLozupone/kokon/pull/36),- closes #5) and the full local HTTP API reference in [http-api.md](http-api.md)- ([#37](https://github.com/GabrieleLozupone/kokon/pull/37), closes #9), both by shrinivas-sn.-+- **0.10.2** (2026-10-23): new image from reference images (also in MCP and the HTTP API), a Prompt+ menu to+ read and change the prompt first, longer Prompt+ prompts, edits never reuse their source's seed, Vulkan decodes+ in 512 px tiles (AMD: seconds instead of 25-70 s), engine+ [kokon-1013-1814814](https://github.com/GabrieleLozupone/stable-diffusion.cpp/releases/tag/kokon-1013-1814814),+ `kokon-mcp --help` ([#36](https://github.com/GabrieleLozupone/kokon/pull/36)) and the full HTTP API reference+ ([#37](https://github.com/GabrieleLozupone/kokon/pull/37)) by shrinivas-sn.- **0.10.1** (2026-10-22): own model files never take a text-only Qwen3 as the text encoder; Linux NVIDIA blankwindow fixed; the progress card fits small windows ([#32](https://github.com/GabrieleLozupone/kokon/issues/32));open the images folder from Settings and the gallery ([#34](https://github.com/GabrieleLozupone/kokon/issues/34));diff --git a/site/index.html b/site/index.htmlindex 8d181f0..08948b5 100644--- a/site/index.html+++ b/site/index.html@@ -43,7 +43,7 @@"url": "https://kokon.app/","downloadUrl": "https://github.com/GabrieleLozupone/kokon/releases/latest","installUrl": "https://github.com/GabrieleLozupone/kokon/releases/latest",- "softwareVersion": "0.10.0",+ "softwareVersion": "0.10.2","releaseNotes": "https://github.com/GabrieleLozupone/kokon/releases/latest","softwareHelp": { "@type": "CreativeWork", "url": "https://kokon.app/prompt-guide/", "name": "Prompt guide for Qwen-Image 2.1" },"applicationCategory": "MultimediaApplication",diff --git a/site/llms.txt b/site/llms.txtindex 5c5eb35..a32527f 100644--- a/site/llms.txt+++ b/site/llms.txt@@ -2,12 +2,12 @@> Kokon is a free, open-source desktop app for editing and creating photos with AI on your own computer, a free, local alternative to Nano Banana and ChatGPT image editing: photos stay on the user's computer and there is no daily limit. Kokon picks and sets up the model for you. You describe the change in plain words ("remove the scooter", "make it a snowy night") and Qwen-Image 2.1 does it locally on your GPU. There is nothing to set up: one installer detects the hardware, downloads and verifies the right model, and picks the resolution and the VRAM/RAM split for every job. Nothing is uploaded.-Key facts (version 0.10.0, October 2026):+Key facts (version 0.10.2, October 2026):- Model: Qwen-Image 2.1 Turbo (official 8-step version, 4-bit GGUF) with Qwen3-VL 8B as the text encoder, run through stable-diffusion.cpp. About 12 GB of models, downloaded once and checked with SHA-256.- Platforms: Windows 10/11 with an NVIDIA GPU (12 GB VRAM or more, CUDA; 8 GB in beta, about 1-megapixel images). AMD and Intel Arc (12 GB+) through Vulkan, beta. Apple Silicon Macs on macOS 15 or newer (Metal), beta: 16 GB works slowly, 32 GB recommended. Linux x86-64 (.deb for Ubuntu/Debian/Mint, .rpm for Fedora; CUDA or Vulkan), beta.- Speed: RTX 5070 Ti about 24 s per edit at 1.5K and 36 s for a new image at native 2K; RTX 5070 12 GB about 12 s per 1-megapixel image; M2 Pro about 4 minutes per image.-- Features: Prompt+ (a few words in any language become a detailed prompt, written by the Qwen3-VL text encoder already loaded, inside Kokon's own build of stable-diffusion.cpp, so no second model), your own Qwen-Image 2.1 LoRAs (beta, tested on Windows), create from text, edit with a sentence, paint an area and Change/Remove/Replace/Add only there, combine several images ("image 1", "image 2"), expand beyond the borders (outpainting), AI upscaling, versions with undo and before/after compare, private mode, works offline after setup.+- Features: Prompt+ (a few words in any language become a detailed prompt, written by the Qwen3-VL text encoder already loaded, inside Kokon's own build of stable-diffusion.cpp, so no second model), your own Qwen-Image 2.1 LoRAs (beta, tested on Windows), create from text, new images from reference photos (a person, a product, a style), edit with a sentence, paint an area and Change/Remove/Replace/Add only there, combine several images ("image 1", "image 2"), expand beyond the borders (outpainting), AI upscaling, versions with undo and before/after compare, private mode, works offline after setup.- For AI agents: built-in MCP server (generate_image, edit_image, inpaint_image, outpaint_image, upscale_image, ...) and a local HTTP API. `npx -y kokon-mcp`; listed in the official MCP Registry as io.github.GabrieleLozupone/kokon.- License: the app is AGPL-3.0. The Qwen-Image 2.1 weights are under the Qwen Research License, which does not allow commercial use.- Not yet: a headless server / Docker image (planned); code signing (Windows SmartScreen and macOS Gatekeeper warn on first run).