comfyUI-Realtime-Lora

Why we included this project

ComfyUI users who train LoRAs will recognize the usual friction: you build a training setup, run it, then hop back into the graph to test the result. This toolkit removes that round trip by handling training, analysis, and selective block loading inside ComfyUI itself, with three backends (Kohya sd-scripts, Musubi Tuner, and AI-Toolkit) behind a single set of nodes that covers SDXL, SD 1.5, FLUX, Z-Image, Qwen Image, and Wan 2.2. SD 1.5 and SDXL runs finish in minutes on a decent GPU, so you can lock down a subject or style and immediately reuse the LoRA in the same workflow. The V2 analyzer and selective loader nodes add per-block strength control, strength scheduling, and support for decomposed formats like LoKR and LoHa, so you can tune exactly how a LoRA affects a generation. For character consistency, style transfer, or Wan video keyframe work, the on-the-fly training and block-level editing save real time.