Combining Multiple LoRAs

Some LoRA-capable models let you stack two adapters at once — e.g. a character LoRA plus a style LoRA. In the LoRA / Integrations modal, put your main LoRA in the LoRA Weights field and a second one in the Extra LoRA field (when the model exposes it). Balance their strengths with the LoRA Scale slider, include every trigger word in your prompt, and keep total influence moderate to avoid oversaturation.

Why Combine LoRAs?

A single LoRA teaches the model one thing — a face, a style, a concept. Combining them lets you build layered results: "my character, in anime style, in a fantasy environment."

Each LoRA operates on different aspects of the generation, so they can complement each other when chosen carefully. A character LoRA modifies the attention layers responsible for facial features and identity, while a style LoRA adjusts the layers that control color palette, texture, and rendering approach.

The key challenge is weight balancing — too much total LoRA influence oversaturates the image and produces artifacts. When the combined weight matrices push activations outside the range the base model expects, you get color blowouts, distorted anatomy, and incoherent compositions.

Stacking Two LoRAs (`Extra LoRA`)

Open the LoRA / Integrations modal on a design and pick a model that exposes both a primary LoRA field and an Extra LoRA field — not every model has one, so if you don't see it, switch the model in the picker at the top.

Put your main adapter in LoRA Weights and the second one in Extra LoRA. Both fields accept the same references: a Replicate owner/model, a Hugging Face huggingface.co/owner/model, a CivitAI civitai.com/models/12345 page, or a direct .safetensors URL.

Use the LoRA Scale slider to set the primary strength; depending on the model, an additional strength control may appear for the extra adapter. A typical setup pairs a trained character LoRA as the primary with a community style LoRA as the Extra LoRA — consistent identity with a stylistic overlay.

You must include trigger words from both LoRAs in your design's prompt. For example, if your character LoRA uses "JDOE" and the style LoRA uses "in the style of FRSTLANE", include both in the prompt text.

Weight Balancing Rules

Keep total combined influence moderate

When the sum of two LoRA strengths gets too high (roughly above 2.0), the accumulated modifications to the base model's attention layers push activations into unstable ranges, producing artifacts. If things look blown out, pull one strength down.

Give the priority LoRA more weight

Decide which LoRA should dominate and weight it higher. A good two-LoRA starting point is roughly 0.7–0.9 for the primary and 0.3–0.5 for the secondary, then adjust from there.

Type-based weight priorities

Character LoRAs usually need higher weight (0.8–1.0) than style LoRAs (0.3–0.5) to maintain facial likeness. If colors look oversaturated or shapes are distorted, reduce weights starting with the style LoRA.

Common Combinations

Character + Style

Face LoRA at 0.9 + style LoRA at 0.4 — the character stays recognizable while the style shifts the aesthetic. This is the most common combination and tends to be the most forgiving.

Character + Clothing

Face LoRA at 0.8 + wardrobe LoRA at 0.7 — both need moderate weight for recognizable results. The character must be identifiable while the clothing concept needs enough weight to appear correctly.

Style + Concept

Style LoRA at 0.5 + concept LoRA at 0.6 — layer an artistic style over a specific scene or object type. Keep the style LoRA lower so it enhances rather than overrides the concept.

Avoid

Combining two conflicting styles (anime + photorealistic) or two different characters at high weights. The LoRAs will compete for the same attention layers and produce incoherent results.

Tips

Always include trigger words from both active LoRAs in your prompt. Missing a trigger word effectively disables that LoRA's contribution.
Test each LoRA individually first to understand its effect before combining. This gives you a baseline so you can tell which LoRA is responsible for any artifacts in the combined output.
If one LoRA overpowers another, reduce its weight rather than increasing the other. Increasing the weaker LoRA raises the total and risks quality degradation.
Both LoRAs must share the design's base architecture — mixing a Flux LoRA with an SDXL LoRA produces broken output because the weight dimensions won't align.
Once a combination works, click Save to store both references and strengths as the design's preset so you don't have to re-enter them.