Skip to content
Let's Comfy

Using LoRA

LoRA (Low-Rank Adaptation) is essentially a “painter’s assistant.”

Not happy with something in the painter’s output? Let LoRA fine-tune it — it can target specific details or shift the overall style and texture.

LoRA files are typically much smaller than checkpoints, and the community has an abundance of them. A LoRA can work standalone or stack with other LoRAs — a very handy little tool.

What You Need

Section titled “What You Need”
  1. Base model (checkpoint): Place it in ComfyUI/models/checkpoints as usual.
  2. LoRA file: Place it in ComfyUI/models/loras (note the plural — don’t misspell).
  3. LoRAs from the web (including LyCORIS, loha, and other variants) work roughly the same way in ComfyUI: load the node, select the file, wire it up.

The common pattern in official examples: base model + a sample LoRA, compared with “no LoRA” vs “with LoRA” on the same prompt — you’ll see clearly: same base model, completely different vibe.

How to Wire It in a Workflow

Section titled “How to Wire It in a Workflow”

The most common node is Load LoRA (or similar):

  • model / clip inputs: Connect from the Checkpoint (or a previous LoRA node).
  • lora_name: Select the file from the loras folder.
  • strength_model: How much this LoRA affects the image model weights — higher means stronger style/character influence; too high causes blurring, artifacts, or over-fitting to a face.
  • strength_clip: How much this LoRA affects CLIP (the text understanding part) — changing this means the model interprets your prompt through the LoRA’s lens.

Output is still model + clip, which passes straight to the prompt encoder and sampler.

Chaining Multiple LoRAs

Section titled “Chaining Multiple LoRAs”

You can chain several Load LoRA nodes in series: the output of one feeds into the input of the next. Order and strength are both variables — pure experimentation. It’s like adding multiple seasonings to a dish: too many will clash.

For a more systematic approach, see the official multiple LoRAs tutorial.

Things to Try When Learning

Section titled “Things to Try When Learning”
  1. Fix the prompt and seed, change only strength_model, and observe the style shift.
  2. Browse Civitai and try a few LoRAs to get a feel (mind the license terms).
  3. Remember: LoRA isn’t magic — if the base model is too mismatched, even a great LoRA will feel awkward.

Further Reading

Section titled “Further Reading”