This ComfyUI node enables direct LoRA training, analysis, and selective loading within the UI. It targets creators seeking rapid LoRA prototyping and fine-grained control over model fine-tuning without leaving their workflow, streamlining the LoRA development and application process.

How It Works

It unifies three training backends (sd-scripts, Musubi Tuner, AI-Toolkit) for a single interface supporting SDXL, SD 1.5, FLUX, Z-Image, Qwen, and Wan models. Advanced analysis tools identify LoRA block-level impact, complemented by selective loaders for per-block control and strength adjustment, offering precise output manipulation.

Quick Start & Requirements

• Installation: Clone into ComfyUI custom_nodes.
• Prerequisites: Python 3.10-3.12 recommended. Training requires separate installation of sd-scripts, Musubi Tuner, or AI-Toolkit backends and specific model downloads. RTX 50-series GPUs need PyTorch 2.7+/CUDA 12.8 via a community installer.
• Links: Backend repos: kohya-ss/sd-scripts, kohya-ss/musubi-tuner, ostris/ai-toolkit. RTX 50-series installer: omgitsgb/ostris-ai-toolkit-50gpu-installer.

Highlighted Details

• Real-time LoRA training and editing directly within ComfyUI.
• Broad model support: SDXL, SD 1.5, FLUX, Z-Image, Qwen, Wan 2.2.
• LoRA Analyzer visualizes block impact (0-100%); Selective Loaders offer per-block control and strength adjustment.
• Claims rapid training times, e.g., ~2 minutes for SD 1.5.

Maintenance & Community

Maintained by Peter Neill (ShootTheSound.com), aiming to democratize LoRA training for creators. Feedback via GitHub issues is welcome. Author accepts support via donations. No specific community channels listed.

Licensing & Compatibility

MIT license, generally permitting commercial use, modification, and distribution.

Limitations & Caveats

Training requires separate backend installations and model downloads. AI-Toolkit on RTX 50-series needs a specialized PyTorch/CUDA setup. Python 3.13 is not recommended. Certain model variants (e.g., Qwen fp8) are unsuitable for training.