How to Speed Up Image Generation with LCM LoRA's (A1111)

Revolutionize your image generation speed by up to 1000%! Welcome to the cutting edge of image synthesis innovation. In this tutorial, we explore the transformative realm of Latent Consistency Models (LCMs), inspired by the success of Latent Diffusion models (LDMs). Fueled by the quest for swifter and more efficient image generation, LCMs bring about a revolution, allowing rapid, high-fidelity sampling with minimal steps on pre-trained LDMs. Dive in with us as we turbocharge your stable diffusion image creation for unparalleled speed and quality.

To speed up your image generation in Stable Diffusion, a single LCM LoRA is all you need, implementable directly within the prompt, just like regular LoRAs. Visit the official Latent Consistency Models Hugging Face Page for access.

LCM LoRA File Pages (Hugging Face)

Both LCM SDXL and LCM SD 1.5 LoRA's are at your disposal:

• LCM SDXL LoRA Files Page
• LCM SD 1.5 LoRa Files Page

From there you can download the "pytorch_lora_weight.safetensors" file

I suggest personalizing the name of the LCM LoRA to match your version. For example, after downloading the LCM SD 1.5 LoRa mentioned above, I changed the file name to "LCM_SD1.5.safetensors," and for the LCM SDXL LoRa, I used "LCM_SDXL.safetensors".

Downloaded LCM LoRA File Location

Similar to downloading regular LoRAs, you'll want to relocate the downloaded file to the LoRa models folder, located at: "stable-diffusion-webui\models\Lora".

Now that our downloads have more personalized names than a secret agent, we're on the express lane to lightning-fast image generation! Let's roll up our sleeves and dive into incorporating these LCM LoRAs into our Stable Diffusion workflow. As we proceed, we'll need to fine-tune certain settings and configurations to ensure our speed approaches lightning-fast levels.

Let's open the stable diffusion web UI and let's see how we can implement our LCM LoRA for blazing fast image generation speeds. Now we need to make sure we select our LoRA, this can be done by:

Adding a LCM LoRA

• Go to the "txt2img" tab within stable diffusion
• Below the prompt section select the "Lora" subtab

• From there select the LCM LoRA you want to use (We will use the LCM_SD1.5 LoRA)

After selecting it it will add the LoRA to your prompt with a weight of 1 like so:

LoRA Models Tutorial Video

Not sure how LoRA's work? Checkout out our Video Tutorial about LoRA Models.

For the LoRA to effectively enhance image generation speed, it's crucial to establish specific settings. Let's kick off with the settings, ensuring strict adherence for optimal generation speed:

Stable Diffusion Settings for Optimal Speed

• Sampling Steps: Set between 2-8 (8 yields the highest quality results).
• CFG Scale: Choose between 1-2 (Opting for 2 includes the negative prompt; CFG Scale of 1 will exclude the negative prompt).

Official LCM LoRA Article

For further details, refer to the official LCM LoRA Article.

As the CFG scale is set lower than usual (around 7), prompts are likely to carry less weight during image generation. To ensure your prompt still influences the generation, consider the following:

• Keep your prompt brief and concise.
• Assign more weight to keywords in your prompt.

As an illustration, the subsequent prompt was employed to generate the examples:

Examples

Let's now explore some recommendations if you encounter challenges in obtaining high-quality images with the LCM LoRAs.

Diverse sampling methods may not always yield optimal image quality. It's recommended to experiment with various methods at each checkpoint for superior image results. In my opinion, Euler, Euler A, or DDIM prove to be the most effective and practical choices.

Poor Quality Images? XYZ Plot!

Before raising concerns about image quality, consider exploring these options or employing an XYZ plot for comparison. Interested in learning how to utilize the XYZ plot to assess checkpoints against sampling methods?

XY Plotting in Stable Diffusion

I hope you're pleased with your selected checkpoint and the swift image generation speeds it provides.

In summary, our exploration into Latent Consistency Models (LCMs) reveals a paradigm shift in image synthesis. Focused on speed and efficiency, LCMs offer a revolutionary approach, enabling swift, high-quality sampling with minimal steps. As we integrate LCM LoRAs into our Stable Diffusion workflow, personalization, setting adjustments, and strategic sampling methods like Euler, Euler A, or DDIM emerge as key factors. Armed with this knowledge, we unlock not just speed but a transformative experience in image generation. Embrace the era of LCMs and elevate your creative journey to new heights.