When using multiple LoRAs to change styles, previous studies have mentioned issues such as the degradation of styles due to overlapping weights. This time, we will apply multiple LoRAs to see the benchmarks and the images produced. We use these LoRAs [LoRa] Shoal / 錦鱗 Concept (With multires noise version)<a href="https://civitai.com/models/79737/lora-shoal-concept-with-multires-noise-version" rel="noopener noreferrer" target="_blank">https://civitai.com/models/79737/lora-shoal-concept-with-multires-noise-version</a> [LoRa] Swatches Neon / 色塊ネオン Concept (With multires noise version)<a href="https://civitai.com/models/83847/lora-swatches-neon-concept-with-multires-noise-version" rel="noopener noreferrer" target="_blank">https://civitai.com/models/83847/lora-swatches-neon-concept-with-multires-noise-version</a> In img2txt, selecting multiple LoRAs will result in a prompt with multiple LoRA specifications mentioned. The following prompt was used. <pre class="ql-syntax" spellcheck="false">fish, neon &lt;lora:underwater3:1&gt; &lt;lora:cbpk:1&gt;
</pre> Finally, two LoRAs are introduced within &lt; &gt;. First, I will run the model with both LoRAs removed and with a specified fixed seed. The base model used was Goofball mix.<a href="https://civitai.com/models/53761?modelVersionId=285757" rel="noopener noreferrer" target="_blank">https://civitai.com/models/53761?modelVersionId=285757</a> without LoRA / RTX4090 / 512x512 / 1.7s<img src="https://assets.blueqat.com/public/uploads/us-east-2:4805ff4b-c3cc-4344-b165-86544c34d0bf/2024/03/20/lora0_0.png"/> with neon LoRA / RTX4090 / 512x512 / 3.3s<img src="https://assets.blueqat.com/public/uploads/us-east-2:4805ff4b-c3cc-4344-b165-86544c34d0bf/2024/03/20/lora2_2.png"/> with underwater LoRA / RTX4090 / 512x512 / 3.5s<img src="https://assets.blueqat.com/public/uploads/us-east-2:4805ff4b-c3cc-4344-b165-86544c34d0bf/2024/03/20/lora1_1.png"/> with neon &amp; underwater LoRA / RTX4090 / 512x512 / 5.1s<img src="https://assets.blueqat.com/public/uploads/us-east-2:4805ff4b-c3cc-4344-b165-86544c34d0bf/2024/03/20/lora3.png"/> How about that? Even with a fixed seed, quite different styles emerged. Each individual LoRA seems to work well, capturing the intended atmosphere. Mixing LoRAs produces another distinct style, but blending completely different styles like underwater and neon can result in a confusing, yet artistic look. It seems interesting to combine LoRAs according to one&#39;s purpose. The computation time increased when multiple LoRAs were applied.

Multiple LoRA with Stable Diffusion on RTX 4090 benchmark

Yuichiro Minato