About this LoRA

This LoRA aims to reproduce a specific amateur, low-quality photo style from the old Grok Imagine model, now available under the "speed" toggle.

To be clear, this isn't meant to capture every look Grok can produce. Everyone has a different experience with it and a different style they like. Grok has a really wide range, even within the narrow category of "amateur low-quality" photos, but there's one particular look that's my favorite. That's what I'm trying to capture here.

Grok Imagine Speed is one of my favorite models of all time. To me, there's still nothing that matches it for truly candid amateur shots, so I wanted to train a LoRA before it's too late. You never know when a model will be removed.

Of course, a LoRA will never be as good as the full model trained by a billion-dollar corporation with a focus on NSFW amateur candid imagery. Nothing else really comes close to that look, but this is my attempt to preserve a small part of it.

Training and dataset

I've done a few training runs over the past few weeks, and this is the first version I consider good enough to release. It's also my first time training a style LoRA, so there was definitely a learning curve.

I built the dataset by hand, generating around 2,000 images with Grok Imagine and writing every prompt myself. I then went through them and picked the images that best matched the style I was going for.

I used Qwen 3.6 to caption the selected images, but literally none of the captions were 100% accurate. I had to manually correct or rewrite a lot of them. From there, I trained the LoRA, revisited the dataset and captions, and repeated the process a few times until I got a result I personally liked.

Known issues

Similar-looking characters

The LoRA tends to lean toward a specific look, not just in terms of style, but also the women it generates.

That doesn't mean it can't produce variety, but if you don't specify the woman's appearance, you'll notice a similar look across generations. I tried to reduce this through the captions, but it's still noticeable, so you'll need to prompt for the differences you want.

Color and hue shifts

You'll probably also notice the colors shifting from warm and reddish to cool and greenish, then back again across generations.

I went back through the dataset, and Grok does produce a lot of images like this, so I guess the LoRA picked up that aspect too. If it bothers you, this version might not be for you.

Mostly indoor photos

Not necessarily an issue, but as the name suggests, around 90% of the dataset consists of indoor photos, mostly in bedrooms, kitchens, living rooms, and bathrooms. It's worth keeping that in mind, since those are the environments where this LoRA works best.

Testing

Unfortunately, testing is what I've spent the least time on so far.

I did around 200 generations to check that it works across a wide variety of poses and doesn't fuck up prompt adherence compared to using no LoRA. It passed that test, which was the most important thing for me (other than the style actually working), but testing beyond that has been limited.

Recommended LoRAs

There are two LoRAs I'd suggest using alongside this one:

- TextFusion Refusal Reduction v2: Almost a necessity. If you're here, I'm pretty sure you already have it.

- Low resolution slider by iamdttla: One of my favorite LoRAs. It further enhances the low-quality look I'm going for. I recommend using it at around 1.5 strength.

Plans for the future

Once I get some feedback, I might train a newer, better version that fixes some of these issues and takes your suggestions into account. Please leave feedback after testing the LoRA, I'd like to hear how it works for you.

If enough people are happy with it and share my love for this specific style, that would definitely motivate me to keep working on it. I might try to improve it anyway, but we'll see.

Usage

• Trigger word: sp33dgr0k, though it might not actually be necessary.

• Recommended strength: 0.7-1.2.

• Recommended sampler/scheduler: er_sde + simple

• Steps: 9

• The prompts can be pretty simple. I'd suggest not prompting for the style at all, just let the LoRAs do the work. Take a look at my prompts for examples.