ComfyUI

Rather than sliders on a form, ComfyUI presents generation as a graph: you drop in nodes for loading checkpoints, encoding prompts, sampling, decoding, masking, and saving, wire them together, and queue the graph to run. Re-running a graph skips nodes whose inputs have not changed, and each output file stores the graph that made it, so loading an old image restores the full setup, seeds included. Native support spans Stable Diffusion 1.5 through SD3.5, Flux, Qwen Image, HiDream, video models such as Wan, LTX-Video, HunyuanVideo, and CogVideoX, plus audio and 3D models, along with LoRAs, ControlNets, and upscalers. Careful memory management and weight offloading let big models run on cards with only a few gigabytes of VRAM, across NVIDIA, AMD, Intel, Apple, and Ascend hardware. Install it as a Windows or macOS desktop app, a portable Windows build, through comfy-cli, or by hand in Python; a local HTTP API turns workflows into backend services. The core is GPL-3.0, and optional paid API nodes for closed models can be disabled with an offline flag. A huge custom-node ecosystem makes it a favorite of power users and studios.