Load LoRA (Downloaded)

Of all the "(Downloaded)" loaders in ComfyUI Model Batch Downloader, this is the one you'll touch most. LoRAs are the ecosystem's default way to add a style, character, or concept on top of a base model - they're small, shareable, and every good workflow stacks two or three. Load LoRA (Downloaded) does what the standard Load LoRA does, except the file comes from the batch downloader's download_result instead of a folder dropdown.

How it sits in the graph

Unlike the other loaders, this one isn't a source - it's a transformer. It takes an existing MODEL and CLIP (typically from a checkpoint or the diffusion/CLIP loaders), applies the LoRA's weights to both, and outputs modified MODEL and CLIP. The chain is: base loader → one or more of these → sampler. Stack several in a row and each one applies on top of the last, which is how you get a character LoRA and a style LoRA in one image.

That in/out shape means it's also a drop-in replacement for the core Load LoRA node in an existing graph - you can swap the file source to the batch downloader without rewiring anything downstream.

Inputs that matter

Five required inputs, and honestly only three of them need your attention:

• model and clip - wire in whatever's upstream (from the checkpoint or split loaders).
• download_result - from the download node.
• id - the manifest ID of the LoRA file (filename stem if unset).
• strength_model and strength_clip - both default to 1.0, range −100 to 100, step 0.01.

The two strengths are where the tuning happens. strength_model controls how much the LoRA changes the diffusion model; strength_clip controls how much it changes the text encoder. For the common case they move together at 1.0. The README's specific tip: for model-only LoRAs, set strength_clip to 0. Some LoRAs train no text-encoder component - typically character/concept LoRAs - and cranking the CLIP side anyway is how you get a LoRA that fights your prompt instead of complementing it. If a LoRA looks "too strong" on the prompt side but perfect on the model side, that slider is the dial.

The range going negative is intentional too - negative strength inverts a LoRA, which is a real (if niche) technique for subtracting a style.

The guard, same as the rest

The loader checks that the id's manifest entry has model_type: loras. File it under the wrong category and you get the pack's clear "requires loras" error instead of a confusing load failure. Keep LoRA downloads in loras and this never bites.

Install and troubleshooting

Shared pack install: ComfyUI Manager search for "ComfyUI Model Batch Downloader", or:

cd ComfyUI/custom_nodes
git clone https://github.com/watarika/ComfyUI-Model-Batch-Downloader.git

then restart. No Python dependencies; aria2c on PATH for the download half.

Failures: an id missing from the manifest (typo or failed download - interrupted ones resume next run), the category-mismatch guard, and the tuning ones that aren't really errors: a LoRA that does nothing at 1.0 strength is usually a LoRA trained for a different base model than your checkpoint, and one that overwhelms the image is usually strength_clip doing more than you wanted. The model-only tip above fixes the most common version of that second one.