Guide to Using Hyper Turbo / Distilled LoRAs

What Are Hyper / Turbo LoRAs?

Hello, this is Mirai. Although I have covered this topic before, there is some new information regarding hyper/turbo/distilled LoRAs that I want to share.

What is a hyper/turbo/distilled LoRA? Many PixAI users use them, but might not know their main function or how they actually work.

Essentially, these LoRAs make your image generation run faster, requiring only a few steps. If you have a low-end GPU (for local setups), this saves time. For PixAI users, it primarily saves credit usage (e.g., only 1200 credits in high priority).

Here are the top three Distilled LoRAs commonly used by PixAI users:

• DMD2
• LCM & Turbomix
• Hyper SDXL

Recommended Parameter Settings

Normal LoRA Parameters

Sampling method: DPM++ 2M Karras
Steps: 20
CFG: 5

Distilled LoRA Parameter Settings

Sampling method: Euler A
Steps: 8
CFG: 1.5

Results: Base Style

Normal

DMD2

LCM & Turbomix

Hyper SDXL

Results: Artist Style

For this part, I tested how distilled LoRAs affect specific artist styles. I used the tag redrop to control the style.

Normal

DMD2

LCM & Turbomix

Hyper SDXL

Discussion

As you can see, we are comparing how the results differ between these three distilled LoRAs.

For normal usage (without prompting a specific style), applying a distilled LoRA noticeably alters the look. Because I used Haruka v2 (the most popular SDXL model in PixAI presets), it has a default 2.5D AI style. Adding a distilled LoRA tends to heavily enforce that 2.5D look.

Remember, since the Illustrious model is fine-tuned on base SDXL (which is originally a realism model), you can see elements of realism creeping in with LCM Turbomix and Hyper SDXL.

However, DMD2 behaves differently. I wonder why it results in much higher saturation compared to the other two.

The second test table shows the most interesting part: using an artist style. In the normal generation steps, the model accurately replicates redrop's vibrant style.

When you try to invoke an artist's style alongside a distilled LoRA, it loses some accuracy. The "flattening" effect of the distillation process is very obvious.

Why Does This Happen?

Distilled LoRAs work by skipping the sampling steps normally needed to refine an image, resulting in faster generation. Because you must set the CFG scale around 1.0~1.5, the AI essentially ignores all negative prompts. This is great for quickly generating decent results without a long wait time (though in PixAI, the main motivation is credit cost).

The downside is that your negative prompts won't work, and stylistic nuances are often lost in trade for speed.

When Should I Use a Distilled LoRA?

For PixAI users running low on credits, using a distilled LoRA is an excellent option. Some users mistakenly believe that distilled LoRAs add more detail to the results (in my experience, they often just "fry" or over-saturate the image). The real advantage is simply that they are much cheaper to generate with.

However, if you are the type of creator who wants to generate images with high style accuracy and intricate prompt adherence, it is not recommended to use these LoRAs due to the trade-offs in their nature.

Anyway, have fun experimenting with distilled LoRAs! You can absolutely use them freely, just be cautious when combining them with other LoRAs, as not all of them are compatible with the distilled process.