ComfyUI gets a LoRA laboratory, not just a loader

What it does

This ComfyUI extension turns the visual node editor into a full LoRA workshop. It wraps three distinct training backends—sd-scripts, Musubi Tuner, and AI-Toolkit—so you can train, analyze, block-edit, merge, and schedule LoRAs for everything from SD 1.5 to Wan 2.2 without switching windows. The analyzer and loader nodes work out of the box; training nodes need you to install the corresponding backend separately.

The interesting bit

The V2 nodes treat LoRAs like modular synthesizers: you can inspect per-block contributions, fade strengths over time with 40-odd schedule presets, and even extract the mathematical difference between two model states into a new, distributable LoRA file. It is essentially model surgery performed with sliders and wire connections.

Key highlights

• Supports eight architectures—SD 1.5, SDXL, FLUX, FLUX Klein 4B/9B, Z-Image, Qwen Image, Qwen Image Edit, and Wan 2.2—through three training backends.
• Block-level control: selectively enable, disable, or scale individual layers, and save the filtered result as a new LoRA.
• Model Diff to LoRA node merges, compresses rank, or selectively “bakes” only the layers you want into a single file.
• Strength scheduling built into the loader: apply keyframed fades, pulses, and bell curves during generation.
• Analysis, selective loading, and model layer editing require no extra backend installation; training nodes do.

Caveats

• Training is not turnkey: you must manually install sd-scripts, Musubi Tuner, or AI-Toolkit, plus download the correct base models and VAEs for each architecture.
• The README notes that RTX 50-series cards need a community PyTorch/CUDA workaround specifically for AI-Toolkit training.

Verdict

Worth installing if you live in ComfyUI and want to iterate on LoRAs fast—especially for SDXL and SD 1.5, which train in minutes. If you only occasionally download premade LoRAs, this is overkill.