EasyLLLite

ControlNet-LLLite is kohya-ss's lighter-weight take on spatial conditioning: instead of a full trainable copy of the model's encoder (what a normal ControlNet is), it injects a small, LoRA-like module at specific attention and MLP points. Same basic idea as ControlNet - condition generation on an edge map, a depth map, a pose skeleton, whatever - but a much smaller file and a lighter footprint at inference. easy LLLiteLoader - labeled EasyLLLite on the canvas - is Easy-Use's wrapper around kohya's original ControlNet-LLLite-ComfyUI node, per the pack's own credit table.

How it works

You give it your model, pick an LLLite model file, and feed in a cond_image - the same kind of conditioning image (canny edges, depth, pose, whatever the specific LLLite model was trained on) you'd feed any ControlNet. The node patches your model with the LLLite module and returns the patched model, ready to sample with normally.

The install gotcha that actually matters

This is the one thing worth getting right before you touch the node at all. Straight from the pack's own README: LLLite model files need to go in ComfyUI\models\controlnet\ - the same default folder ComfyUI already uses for regular ControlNet models, not a separate LLLite-specific folder - and the filenames need to stay exactly as downloaded. Rename the file and the node won't be able to read it. If model_name's dropdown is empty or your model isn't showing up, this is almost always why: wrong folder, or a renamed file.

The inputs and outputs that matter

• model - the model to patch. Required.
• model_name - the LLLite model file, populated from your controlnet folder.
• cond_image - the conditioning image, matching whatever type the specific LLLite model expects.
• strength (default 1, range 0-10) - how strongly the conditioning is applied.
• steps (default 0, range 0-200) - an internal scheduling value the loader uses; leave it at the default unless a specific LLLite model's documentation tells you otherwise.
• start_percent / end_percent (default 0, range 0-100) - the window during which the conditioning is active, as a percentage. Worth noting explicitly: most other percent-style controls elsewhere in this pack (like the downscale sampler's start_percent/end_percent) use a 0-1 fraction, not 0-100. This node uses whole percentages - check which convention you're looking at before you type "0.35" here and get nearly nothing.

Output: the patched MODEL, wired straight into your sampler.

Installing it

Ships with the base pack for the node itself. ComfyUI Manager: search ComfyUI Easy Use, install, restart. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/yolain/ComfyUI-Easy-Use

then install.bat on Windows or pip install -r requirements.txt, restart. The LLLite model weights are a separate download - get them from wherever the specific LLLite model you want is published, and place them in ComfyUI/models/controlnet/ with their original filenames, per the gotcha above.

Common issues & troubleshooting

model_name dropdown is empty. Model file isn't in ComfyUI/models/controlnet/, or it's there under a renamed filename. Fix the folder/filename and restart ComfyUI so it re-scans.

Conditioning barely seems to do anything. Check start_percent/end_percent first - since this node uses 0-100 rather than 0-1, a value that "looks small" like 0.5 is functionally almost off. Also check strength; LLLite models can need a different strength range than a standard ControlNet you're used to, since it's a different mechanism (a LoRA-like patch rather than a full encoder copy) and the two don't necessarily calibrate the same.

Wrong conditioning image type for the model. LLLite models are trained on a specific conditioning type each - a canny-trained LLLite model fed a depth map (or vice versa) will produce nonsense results. Match the cond_image you generate to whatever type the specific LLLite model file was actually trained on.