Mastering ComfyUI — Part 4: LoRAs & Model Customization

NOTE

Goal for this part: understand what a LoRA actually is, how to add one to a workflow correctly, and how to avoid the most common ways they go wrong.

Previous: Part 3 — Building Your First Workflow

What a LoRA actually is

A full checkpoint is billions of parameters — retraining one from scratch to teach it a new style or concept takes serious compute. A LoRA (Low-Rank Adaptation) is a small file — typically tens to a few hundred megabytes, compared to several gigabytes for a full checkpoint — that captures just the adjustment needed to nudge an existing checkpoint toward a specific style or concept, without touching the base model itself.

Think of the checkpoint as the foundation and a LoRA as a filter applied on top: same underlying model, but pushed in a specific direction. This is why the LoRA ecosystem is so large — it's far cheaper to train and share a small LoRA for a specific style than a fresh multi-gigabyte checkpoint.

Adding a LoRA to a workflow

Search and place "Load LoRA (Model and CLIP)". It sits between Load Checkpoint and everything downstream:

Load Checkpoint's MODEL → LoRA node's model input
Load Checkpoint's CLIP → LoRA node's clip input
The LoRA node's own MODEL and CLIP outputs then go everywhere your CLIP Text Encode and KSampler nodes were previously connected directly to Load Checkpoint

This insertion point matters: a LoRA modifies both the diffusion model's behavior and how prompts get interpreted (since some LoRAs also teach new trigger words), so it needs to sit before both the sampler and the text encoding.

Strength — the setting that actually matters

The LoRA node has a strength_model and strength_clip value, both typically defaulting to 1.0. This controls how strongly the LoRA's influence is applied:

Too low (below ~0.5) and you often won't see the LoRA's effect at all
Too high (much above ~1.2) and you risk the LoRA overpowering the base checkpoint — artifacts, oversaturation, or the LoRA's specific training quirks bleeding into everything regardless of your prompt
1.0 is a reasonable starting point for most LoRAs — adjust from there based on what you actually see, not a formula

TIP

Read the LoRA's own documentation first. Most LoRAs downloaded from community model sites come with recommended strength ranges and, critically, trigger words — specific tokens you need in your prompt for the LoRA to actually activate. Skipping the trigger word is one of the most common reasons a LoRA appears to "do nothing."

Stacking multiple LoRAs

You can chain multiple Load LoRA nodes in sequence — each one's MODEL/CLIP output feeds the next one's input, with the final node's outputs going to your prompt/sampler nodes as usual. Stacking works, but effects compound and can conflict: two LoRAs both trying to push style in different directions will fight each other. Start with one LoRA at a time, confirm it does what you expect, then add the next rather than stacking three or four blind and troubleshooting all at once.

Where LoRAs actually live

Place downloaded LoRA files in your models/loras/ folder — ComfyUI scans it the same way it scans checkpoints/, and a restart or refresh is needed for newly-added files to show up in the dropdown.

NOTE

A note on sourcing models. Community model repositories host a huge range of checkpoints and LoRAs, and content quality and appropriateness vary widely — many such sites mix general-purpose models with adult content. Use the site's content filters, stick to models whose purpose and licensing are clearly stated, and check what a model was actually trained for before downloading it.

Troubleshooting

BUG

LoRA doesn't seem to change anything. Check the trigger word first — most LoRAs need a specific phrase in your prompt to activate, and simply loading the file isn't enough on its own. Second most common cause: strength_model / strength_clip set too low.

BUG

Output looks distorted or oversaturated after adding a LoRA. Strength is likely too high. Drop both strength values toward 0.6–0.8 and re-test before assuming the LoRA itself is broken.

BUG

LoRA works but breaks multi-subject scenes that worked fine without it. Some LoRAs are trained on single-subject data and can worsen the same attribute-bleeding problems covered in Part 5. If a scene that worked cleanly starts merging subjects again after adding a LoRA, the LoRA itself is a likely suspect — test with it removed to confirm before troubleshooting the conditioning setup.

Next: Part 5 — Advanced Conditioning, covering regional conditioning, inpainting compositing, and ControlNet.

Previous: Part 3 — Building Your First Workflow