Combining Multiple LoRAs

You can stack multiple LoRAs to combine a character's face with an art style, a specific outfit, and a scene type — all in one generation. On Replicate, use extra_lora/extra_lora_scale for dual-LoRA setups, or lucataco/flux-dev-multi-lora for up to 20 simultaneous LoRAs. Keep total combined weights under 2.0 to avoid quality degradation.

Why Combine LoRAs?

A single LoRA teaches the model one thing — a face, a style, a concept. Combining them lets you build layered prompts: "my character, in anime style, wearing a specific outfit, 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.

Method 1: Dual LoRA (`extra_lora`)

Available on Replicate-trained fine-tunes, this is the simplest way to combine two LoRAs. Your primary LoRA is the trained model itself, controlled by lora_scale (default 1.0, range −1.0 to 2.0).

Add a second LoRA via the extra_lora parameter — it accepts HuggingFace repo IDs (e.g., alvdansen/frosting_lane_flux), CivitAI URLs, or direct .safetensors URLs. Control its strength with extra_lora_scale (default 1.0, range −1.0 to 2.0).

You must include trigger words from both LoRAs in your 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.

A typical use case: your trained character LoRA as the primary model + a community style LoRA from HuggingFace as the extra_lora. This gives you consistent character identity with a stylistic overlay.

Method 2: Multi-LoRA Model

The lucataco/flux-dev-multi-lora model on Replicate supports up to 20 simultaneous LoRAs. Pass an array of LoRA references via the hf_loras parameter, where each entry is a HuggingFace repo ID or URL.

Each LoRA in the array is applied with equal weight by default. You can control individual weights via the lora_scales parameter, passing a comma-separated list of floats corresponding to each LoRA in order.

This approach is excellent for experimentation and complex compositions where you need more than two LoRAs. All LoRAs must be compatible with the same base model — in this case, Flux. Mixing a Flux LoRA with an SDXL LoRA will produce broken output because the weight dimensions won't align.

Weight Balancing Rules

Keep total combined weights under 2.0

This is the fundamental rule. When the sum of all LoRA weights exceeds 2.0, the accumulated modifications to the base model's attention layers push activations into unstable ranges, producing artifacts.

Starting points by count

When combining 2 LoRAs, start at 0.7–0.8 each. When combining 3+, start at 0.5 each and adjust. Give higher weight to the LoRA you want most prominent in the result.

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 all 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 weight and risks quality degradation.
Negative LoRA weights (e.g., −0.5) can be used to actively suppress a LoRA's effect — useful for undoing unwanted stylistic shifts while keeping other LoRAs active.
Save working combinations as presets in the Integrations modal so you don't have to re-enter URLs each time.