Comfy Workbench — An Unofficial Android Client for ComfyUI

⚡ The Gist in 5 Seconds - What it is: Comfy Workbench is a native Android app (Kotlin + Jetpack Compose) that runs ComfyUI workflows already built on your computer, right from your phone. - Availability: APK in Releases, Android 8.0+, universal build (arm64-v8a / armeabi-v7a / x86 / x86_64); connects to your own ComfyUI instance via HTTP + WebSocket. - Limitation: it's not a node editor — only API-format JSON is accepted, and video files as input are not supported. ### 🔍 What Was Found The Mimikkoro1/comfy-workbench-android project is an unofficial client under GPL-3.0, not affiliated with Comfy Org. Here's how it works: on your computer, a workflow is exported from ComfyUI in API format, then imported into the app, where its parameters are turned into a few fields — prompt, seed, size, reference. Prompts are recognized on CLIPTextEncode nodes and TextEncode-type nodes, including Qwen. There are 11 built-in size presets, including 9:16. The connection goes directly over HTTP + WebSocket with real-time progress and no cloud middleman — prompts, parameters, and images travel only between your phone and your machine. Features include: a prompt library with categories, random selection, and editing; a gallery with unified viewing of images and videos, where results stay inside the app and are not automatically written to the system album; generation and downloads run through a foreground service, so they don't stop when the screen is locked, with progress shown in a notification; and a queue with the ability to interrupt the current task. The UI is Material 3, in Chinese and English, following the system language by default. Update checks ping GitHub no more than once a day, there's no auto-download, and no data is collected. The limitations are stated plainly: you can't connect nodes or build new graphs on the phone, and the default file (nodes / links) won't work — you need Export (API). Video input isn't supported, only a reference image for image-to-video. The control panel is built from literal parameters: already-connected inputs are not shown. Custom sizes are rounded down to a multiple of 16, with a minimum of 64. For remote access, there's an address field and HTTP Basic Auth behind Caddy / Nginx — on a 401, login and password fields appear. Tunneling options include (Oray Pgyer), public IPv6 + DDNS, Tailscale, or frp /; opening port 8188 publicly is not recommended — on untrusted networks, use only a VPN or tunnel. ### 💡 Why It Matters The point of the project is freedom from being tied to your computer. Once a workflow is dialed in, the phone becomes a remote control: the same t2i, i2i, i2v, and prompt inversion (Florence-2 caption) can be launched from the couch, while the graph structure stays on the desktop. It's a deliberate trade-off: you lose the node editor, but you gain one tap instead of fiddling with the web interface. The docs/example-workflows/ folder contains four ready-made examples — kr2turbo_t2i, kr2turbo_i2i, wan22_i2v_4step, and florence2_caption — which you can use to check the format by swapping in your own model names. ### 🧩 Context The project is positioned as personal, "tailored...

🔗 Read on github.com