clark-labs-inc/comfyui-clarkairsana

ComfyUI-ClarkAirSana is a plugin for ComfyUI, the popular node-based interface for AI image generation. It adds support for a model called Clark Air Sana 1.6B, which generates images from text descriptions. The defining feature of this model is that its weights are stored in an extremely compact format called ternary quantization, roughly 1.58 bits per value instead of the usual 16 or 32, which significantly reduces memory usage while keeping image quality practical. The actual computation uses a library called GemLite with INT2 CUDA kernels, meaning it runs on NVIDIA GPU hardware directly. Installing the plugin takes three steps. You install the ComfyUI-ClarkAirSana node pack through ComfyUI Manager, download a 495 MB model file to a specific folder, and drag a provided workflow file onto the ComfyUI canvas. The Gemma text encoder (which converts your text prompts into a form the model understands) and the DC-AE image decoder download themselves automatically on first use. Total download size is about 3.2 GB. The plugin adds four custom nodes to ComfyUI: a model loader, a text encoder, an image decoder, and a latent placeholder. These connect together through the standard KSampler node, so the workflow fits into the same mental model as other ComfyUI setups. The provided example workflow handles all the wiring. There are a few practical requirements to know. The plugin needs an NVIDIA GPU and runs best on Linux or Windows Subsystem for Linux (WSL2). Running on native Windows is possible but may not work reliably depending on the setup. Once all components have been downloaded once, the entire workflow runs offline with no network access needed.