waldbeat – Mega LoRA Loader

waldbeat – Mega LoRA Loader for ComfyUI

Two LoRA loader nodes by waldbeat, built because the Power Lora Loader and its ⓘ info button stopped working with the new ComfyUI frontend (Nodes 2.0):

| Node | Platform | Output | |---|---|---| | Mega LoRA Loader (MLX) | macOS · Apple Silicon (MLX) | lora_stack for the Krea 2 MLX pipeline | | Mega LoRA Loader (Torch) | Windows · Linux · macOS (CUDA / MPS / CPU) | MODEL |

Both share the same interface and the same info dialog. Works in Nodes 2.0 and the classic frontend.

Features

• Loads any number of LoRAs in one stack – one row per LoRA: on/off, LoRA, ◀ strength ▶, ⓘ, ✕
• Search field in the LoRA picker (multiple words, ↑ ↓ Enter, Esc)
• No re-calculation of the model: LoRAs run in bypass mode – the base weights are never patched
• No waiting time after changing a strength or switching a LoRA
• Saves memory: no backup copy of the model weights is kept (regular loaders store the original weights to undo the patch)
• Shows trigger words directly (click to copy) – or "no trigger needed"
• ⓘ info dialog: recommended strength (incl. strength hints found in the description), trigger words, example images, description, version notes
• Info lookup is automatic: local first, then Civitai
  1. LoRA Manager metadata (<lora>.metadata.json) – instant, offline
  2. own cache (<lora>.lsb-info.json) – instant, offline
  3. Civitai by SHA256 hash – fetched once, then cached
• Button to reload the info from Civitai manually (e.g. after the author updated the page)
• Link to the LoRA's Civitai page

Mega LoRA Loader (MLX) – macOS

For Krea 2 running natively on Apple's MLX framework (no PyTorch involved in sampling).

Requirements

• Mac with Apple Silicon (M1 or newer)
• ComfyUI-Krea2-MLX installed
• Connect lora_stack to the lora_stack input of Krea2 Generate (MLX)

Notes

• Output type is KREA2_LORASTACK (list of path + strength), compatible with the Krea2 LoRA (MLX) node – the optional lora_stack input lets you chain it with other LoRA nodes.
• MLX applies LoRAs unfused, so each active LoRA adds a few seconds per image – depending on the LoRA's size. For LoRAs you always use, baking them into the model is faster.
• Negative strengths (slider LoRAs) are supported.

Mega LoRA Loader (Torch) – Windows / Linux / macOS

Standard ComfyUI node: model in → model out. Replaces a chain of LoRA loaders.

Requirements

• Current ComfyUI with bypass LoRA support (comfy.sd.load_bypass_lora_for_models)
• Nothing else – no extra Python packages

Notes

• Model only: the LoRA is applied to the diffusion model, not to the text encoder (CLIP). Most modern LoRAs (e.g. Krea 2, trained with ai-toolkit / diffusion-pipe) contain no text-encoder weights. Connect your CLIP directly from the loader to the text encoder.
• Loaded LoRA files are cached by the node, so switching back to a LoRA costs nothing.

Installation

cd ComfyUI/custom_nodes
git clone https://github.com/waldbeat/ComfyUI-waldbeat

Restart ComfyUI and hard-reload the browser (Cmd/Ctrl + Shift + R). Search for "Mega LoRA Loader".

Optional: set the environment variable CIVITAI_API_KEY for Civitai lookups of restricted models.

Thanks

• rgthree – rgthree-comfy (MIT). The Power Lora Loader and its info dialog were the inspiration for this node. The Civitai lookup by file hash follows his idea.
• willmiao – ComfyUI-Lora-Manager. Its metadata files are used as the first info source.
• Cthomasdesign – ComfyUI-Krea2-MLX, and avlp12 for the MLX port of Krea 2.
• The ComfyUI team – for the bypass LoRA implementation this node is built on.

License

MIT