How to Use LoRAs in ComfyUI: Loader Node, Strength Settings & Stacking

LoRAs—Low-Rank Adaptation files—are compact model add-ons (typically 50–500MB) that teach ComfyUI to generate specific artistic styles, reproduce characters consistently, or handle concepts your base model doesn’t natively support. Using LoRAs in ComfyUI is straightforward: download a .safetensors file, place it in the models/loras/ folder, and wire a LoraLoader node into your workflow. The real skill lies in tuning strength values and understanding how to stack multiple LoRAs without creating visual chaos. This guide covers everything from installation to advanced stacking techniques.

Quick Reference: LoRAs at a Glance

What LoRAs Actually Are

LoRA stands for Low-Rank Adaptation, a fine-tuning technique that modifies only a small portion of a model’s weights—specifically, the deltas that describe a style, concept, or character—rather than retraining the entire model from scratch. The practical result is a compact file that “teaches” the model something new without bloating your hard drive. When loaded, the LoRA is active and changes the model’s behavior; when removed, those changes vanish entirely.

You’ll encounter three main types on Civitai:

• Style LoRAs (50–150MB): Mimic an artist’s brushstroke, visual movement, or color palette. Apply them to any subject and watch the aesthetic shift.
• Character LoRAs (100–300MB): Reproduce a specific character consistently across different prompts and scenes.
• Concept LoRAs (50–200MB): Add objects, poses, or situations the base model doesn’t handle well natively.

All three work identically in ComfyUI LoraLoader—the difference is mainly the recommended strength value and training intensity.

💡 Tip: LoRAs are lightweight adapters that add specific knowledge to your model without replacing it. They’re reversible, swappable, and the fastest way to customize output.

Finding and Downloading LoRAs

Civitai (civitai.com) is where you’ll find tens of thousands of LoRAs, complete with preview images and user comments. Before downloading, filter by base model—this is critical. A LoRA trained for SD 1.5 produces visual garbage with SDXL, and one built for SDXL fails with Flux. There’s no error message; you just get noise.

To filter on Civitai:

1. Navigate to Models
2. Select Type → LORA
3. Expand the Base Model filter and choose your model (SD 1.5 / SDXL 1.0 / Flux.1 D or Flux.1 S)
4. Sort by Most Downloaded or Highest Rated to find proven LoRAs

⚠️ Important: Always download the .safetensors file. It’s the format ComfyUI loads natively and the only format that cannot contain executable code. Avoid .ckpt files unless the LoRA doesn’t exist in any other format—.ckpt files can embed malicious code.

Installing LoRAs in ComfyUI

Copy the .safetensors file directly into ComfyUI/models/loras/. No subfolders are required, though you can organize them if you prefer—ComfyUI finds them either way. No restart needed. The file appears in the LoraLoader node’s dropdown the next time you open it. If it doesn’t show up, refresh the browser (F5).

The LoraLoader Node: Setup and Wiring

Add the LoraLoader node by double-clicking the canvas, typing “lora” in the search box, and selecting LoraLoader. The node has five fields:

• model (input): Receives MODEL from CheckpointLoader or another LoraLoader
• clip (input): Receives CLIP from CheckpointLoader or another LoraLoader
• lora_name (selector): Picks the LoRA file from models/loras/
• strength_model (number): Controls how much the LoRA affects image generation (range –100 to 100)
• strength_clip (number): Controls how much it affects prompt interpretation (range –100 to 100)

Wiring in a standard workflow:

1. CheckpointLoaderSimple’s MODEL output → LoraLoader’s model input
2. LoraLoader’s MODEL output → KSampler’s model input
3. CheckpointLoaderSimple’s CLIP output → LoraLoader’s clip input
4. LoraLoader’s CLIP output → CLIPTextEncode’s clip input

The LoraLoader receives the unmodified model and CLIP, applies the LoRA, and passes the modified versions downstream. This chaining approach is essential for stacking multiple LoRAs.

Strength Values: The Critical Setting

The technical range is –100 to 100, but in practice you’ll rarely leave the –2 to 2 range. What works best depends on your base model and the type of LoRA you’re loading.

SD 1.5 and SDXL

For a new style LoRA, start at 0.8 and adjust ±0.1 to find your balance. Negative values subtract that style—useful to push the image away from certain traits. LoRA strength ComfyUI is model-dependent, so always check the LoRA’s Civitai description for author recommendations.

Flux and Video Models

Flux’s architecture responds differently to LoRAs and requires higher values:

Video models like Wan 2.1 typically use values in the 1.0–1.5 range. Always check the LoRA’s Civitai description—quality LoRAs list the author’s recommended strength.

strength_model vs. strength_clip

In most cases, keep both values the same. strength_model controls how the LoRA affects the diffusion process (the image itself), while strength_clip controls how it affects prompt text interpretation. If the LoRA distorts the prompt too much and text stops working well, lower strength_clip to 0.6–0.7 while keeping strength_model higher.

📌 Keep in mind: When stacking multiple LoRAs, these values compound. Start conservative and adjust from there.

Stacking Multiple LoRAs: Chaining and Tuning

Stack LoRAs ComfyUI by chaining LoraLoader nodes: each node receives the MODEL and CLIP from the previous one and applies its own LoRA on top. The critical rule is to lower the strengths when stacking. If you’d normally use 0.8 with a single LoRA, try 0.5–0.6 each with two—influence accumulates, and high values become unstable together.

Combinations that work well:

• Style + Character: One LoRA defines the palette or brushstroke; the other fixes character identity. Try strength 0.6 + 0.7.
• Style + Lighting: An artistic style LoRA plus a cinematic lighting one. Try strength 0.7 + 0.5.
• Two Styles: Blend of two visual styles. Start with strength 0.5 + 0.5.

Wiring multiple LoRAs:

1. CheckpointLoaderSimple → LoraLoader #1 (model & clip inputs)
2. LoraLoader #1 outputs → LoraLoader #2 inputs
3. LoraLoader #2 outputs → LoraLoader #3 inputs (if using three)
4. Final LoraLoader outputs → KSampler and CLIPTextEncode

Beyond three active LoRAs, results become unpredictable. Use only for deliberate experimentation. Each additional LoRA compounds the influence, so reduce individual strengths progressively—think of it like adding salt to a dish. One pinch is fine; five pinches ruins it.

Negative Strength Values

A negative strength_model doesn’t undo the LoRA; it applies the inverse effect. Practical uses:

• Subtracting a style: If your base model tends to generate faces with an unwanted style (e.g., too anime), that same style LoRA at strength –0.4 can reduce the tendency.
• Fine-tuning combinations: If a character LoRA introduces unwanted traits, a partial negative value can compensate.
• Controlled experimentation: Small negative values (–0.2 to –0.5) give predictable results; very negative values (–1.0 or beyond) produce unpredictable artifacts.

Negative values are rarely used in day-to-day work—more a fine-tuning tool than a habitual workflow element. They’re most useful when you’ve stacked multiple LoRAs and need to dial back one specific influence.

Troubleshooting Common Issues

LoRA Not in the Dropdown List

Check these in order:

1. File location: Verify it’s in ComfyUI/models/loras/, not a parent folder or the checkpoints directory.
2. Browser refresh: Press F5. ComfyUI doesn’t always detect new files in real time.
3. File extension: Must be .safetensors, .pt, or .ckpt. Other formats don’t show.
4. Corrupted download: File size should match what Civitai lists. If it’s 0KB or suspiciously small, redownload.

Results Are Inconsistent or Noisy

• Wrong base model: An SD 1.5 LoRA with SDXL produces noise. Verify the LoRA’s base model on Civitai against your checkpoint.
• Strength too high: Lower to 0.5 and generate several images. If they improve, raise in 0.1 increments.
• LoRA conflict: If using several, disable them one at a time to identify the culprit.
• Fixed seed: Try a random seed to see if the issue is the LoRA or that specific prompt+seed combination.

LoRA Changes Style Too Much Even at Low Strength

Some LoRAs are trained aggressively. Options:

• Lower strength_model to 0.2–0.3 while keeping strength_clip at 0.6 so the prompt keeps working.
• Use an explicit negative prompt to counter unwanted traits.
• Look for another version of the same LoRA—Civitai often has versions with different training intensity.

”LoRA Not Found” When Opening a Saved Workflow

The workflow saves the LoRA’s filename. If you deleted or renamed the file, ComfyUI can’t find it. Fix by updating the lora_name field in the LoraLoader node to select the correct file.

FAQ

Q: Where do LoRA files go in ComfyUI?
A: LoRA files (.safetensors, .pt or .ckpt) go in the ComfyUI/models/loras/ folder. Once copied there, they appear automatically in the LoraLoader node’s dropdown menu without needing to restart ComfyUI.

Q: What strength value should I use for a LoRA?
A: For SD 1.5 and SDXL, use between 0.5 and 1.0. For Flux and video models like Wan, between 1.0 and 2.0. Always start at 0.8 and adjust: above 1.2 on SD/SDXL the LoRA tends to dominate too much and loses coherence.

Q: Can I use multiple LoRAs at once in ComfyUI?
A: Yes. Chain multiple LoraLoader nodes by connecting each one’s MODEL and CLIP output to the next one’s input. There’s no technical limit, though with more than 3 active LoRAs the result can become unpredictable. Keep strengths lower when stacking.

Q: Why doesn’t my LoRA show up in ComfyUI’s list?
A: The most common cause is the file not being in the right folder (ComfyUI/models/loras/) or having an unsupported extension. Verify the path, check the file isn’t corrupted, and refresh the page. If using ComfyUI Manager, check there isn’t an alternate directory configured.

Keep Reading

If you’re still getting familiar with the node graph, our What Is ComfyUI primer covers the basics of wiring nodes together. And once your LoRA setup is dialed in, understanding how KSampler’s steps, CFG and sampler settings work will help you get more consistent results across generations.

🏆 Our Recommendation

If you’re just starting with LoRAs: Download a single style or character LoRA that matches your base model, place it in models/loras/, add a LoraLoader node, set strength to 0.8, and experiment with ±0.1 adjustments. This teaches you how strength affects output without overwhelming complexity.

If you’re combining styles: Stack two LoRAs (one style, one character or lighting) at reduced strength (0.6 + 0.6 for SD/SDXL, 1.2 + 1.2 for Flux). Test one at a time first to isolate their individual effects, then combine.

If you’re troubleshooting: Always verify base model compatibility first—it’s the #1 cause of noise. Then lower strength incrementally. If results still don’t improve, try a different LoRA version or check the Civitai community comments for reported issues.