LoRA Style Mixer

Stable Diffusion LoRA style mixer

LoRAs let you bolt a style, character or concept onto a base Stable Diffusion model — but stacking several means juggling strengths and <lora:...> tags by hand. This mixer lets you add multiple LoRAs with strength sliders, watch the total stacked weight, and copy correctly formatted tags ready to paste.

How LoRA syntax works

In AUTOMATIC1111, Forge and ComfyUI you activate a LoRA by adding a tag to your prompt:

<lora:filename:strength>

filename is the LoRA file in your models/Lora folder (no extension) and strength scales how hard it is applied — 1.0 is full, 0.7 is a common restrained default. Multiple LoRAs simply sit side by side:

a knight in armor <lora:epic_style:0.7> <lora:detail_boost:0.4>

Their effects add up, so two strong LoRAs can overpower the base model and each other. The mixer sums your strengths and flags when the stack gets heavy so you can dial it back before generating.

What LoRA strength actually does

LoRA (Low-Rank Adaptation) works by modifying the model’s weight matrices at inference time. The strength value scales how much each modification is applied:

• 0.0 — the LoRA has no effect; you are generating with the base model alone
• 0.5 — a gentle blend, useful when a style LoRA is too dominant at full strength
• 0.7–0.9 — the typical working range for most style and character LoRAs
• 1.0 — full application of everything the LoRA learned
• >1.0 — overshooting, which can work for some artistic effects but often distorts anatomy or causes colour artefacts

Style LoRAs (which modify the visual aesthetic of everything in the image) typically need lower strengths than character LoRAs (which need to reproduce a specific face or outfit). Detail-boosting LoRAs often work well at 0.3–0.5 stacked on top of a base.

Stacking mechanics and total weight

When you stack multiple LoRAs, their modifications stack in the attention and MLP layers. There is no hard ceiling, but experience shows that a combined effective weight above roughly 1.5–2.0 across style-affecting LoRAs produces interference patterns: the model starts pulling in conflicting directions and coherence breaks down. The mixer totals your weights so you can see when you are approaching that range.

A practical stacking strategy: use one dominant LoRA at 0.7–0.9 for the primary style, and any supporting LoRAs (detail, texture, lighting enhancement) at 0.3–0.5. Keep the total below 1.5 for the first test batch, then raise individual weights selectively if something is underrepresented.

Tips for mixing LoRAs

• Start low. Set each LoRA to 0.6–0.8 and raise only the one that needs more presence.
• Watch the total. A combined weight above ~1.5 across style LoRAs usually muddies results.
• Match the base model. An SD 1.5 LoRA will not work on an SDXL checkpoint — keep the families consistent.
• Use trigger words. Many LoRAs need their trigger word in the text prompt as well as the tag; check the LoRA’s model card.
• Negative LoRA strength. Some UIs support negative strength values (for example, -0.5), which subtracts the LoRA’s learned features rather than adding them — useful for steering away from a particular style.