UPDATE: MOSS V2 - Base model is the way to go

I had a major breakthrough, and it's not with LoRA training. It's with generating the images.

The big discovery

I realized the MOSS base model is much, much better at realistic images than SFT. SFT is just okay. Base is insane. I'm willing to say it's better than Krea2 out of the box, and probably better than Krea2 with many of its LoRAs.

The trick is the ER_SDE sampler node. Please use this workflow.

My settings

• Sampler: ER_SDE (use the node!)

• Steps: 28

• CFG: 3.0

• Uncond skip layers: 22, 24 (this one is extremely important)

Feel free to play around with steps and CFG.

Things to experiment with

• Eta: Lower values (down to 0.1) can sometimes give interesting results.

• S noise: Between 1.05 and 1.1 can slightly change the image.

With all of these together on the base model, the results are fantastic.

UPDATE: Krea2 V2 is here

As promised, I improved the dataset. It's no longer the old crappy low-quality style, though. That look is much harder to get out of Grok Imagine nowadays, with the limits, moderation, and all the other issues it has now.

The new style

While working on this, I noticed the old Grok Imagine model also produces a different style that I like just as much. I'm guessing more people prefer it anyway, so I decided to train a LoRA for it too.

What's better

Most of the old rules still apply, but you can now do a lot more than indoor shots. Outdoors works too: beaches, abandoned warehouses, construction zones, backyards, and so on.

Basically, it's a big improvement in every way.

Usage (KREA 2)

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

• Recommended strength: 0.7-1.1.

• 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.

Recommended LoRAs (KREA 2)

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.

Early access and price

To cover some of my RunPod costs, I'm doing a very short early access with a small price.

I'm not a big fan of charging for LoRAs, so I'll only do this occasionally and at a very affordable price (never above 250 Buzz). Early access lasts 3 days.

UPDATE: MOSS version

I made a version of this LoRA for MOSS, which was released about 2 weeks ago. It's one of the best models out there for styles and aesthetics, and it's absolutely amazing. The catch is that when it came out, nobody really knew how to train LoRAs for it yet. I tried quite a few times and failed miserably. A few days ago I finally got a first decent result that I consider a "release candidate"

It's not perfect yet, because I'm still using the same old dataset I used for Krea. But when it works, it works really well.

Important notes (MOSS version only)

- Use low quality candid smartphone photo, compression artifacts, low dynamic range, dull unsaturated colors, digital noise, JPEG compression, sharpening halos, imperfect white balance in your prompt (or something similar). It's very important. Without it, the results won't be as good.

- MOSS is even more sensitive to my limited, low-variety dataset than Krea was. If your result doesn't look much different from the baseline without any LoRAs, that's probably why.

What it works for

I'll say it again: this only works on dull, low-quality indoor candids. It needs to look dull and boring, borderline depressing. It doesn't like colors or saturation.

Yeah, I know it's a very narrow use case, but that's just the way it is right now. I'm really happy with how it looks when it does work.

Base and SFT

I trained it on both the base and SFT versions. Both work well.

I'll upload the base version after more testing. Turns out Base isn't as good as I initialyl expected so I probably won't be releasing that. I'll focus on training the next version instead.

For SFT, I'm releasing 2 checkpoints: epoch 25 and epoch 65.

Testing LoRAs on MOSS is really hard. The results change a lot from seed to seed. I did hundreds of gens on each epoch and still couldn't find one checkpoint that was clearly the best.

So I picked two:

- Epoch 25: The first checkpoint that worked well. Probably best one.

- Epoch 65: A much later one. I picked it kind of randomly, but I feel like it works well.

Try both and use the one you like more.

Krea2 V1

This LoRA recreates one specific amateur, low-quality photo style from the old Grok Imagine model (now under the "speed" toggle). Grok can do a huge range of looks, but this one is my favorite, and I wanted to preserve a small part of it before the model goes away.

I built the dataset by hand: ~2,000 images generated with Grok Imagine, all prompted by me. I picked the best ones, then captioned them with Qwen 3.6 and fixed a lot of the captions manually.

Known issues:

• It tends to generate similar-looking women unless you describe their appearance.

• Colors shift between warm/reddish and cool/greenish, because Grok's own images do that and the LoRA learned it.

• About 90% of the dataset is indoor photos (bedroom, kitchen, living room, bathroom), so it works best there.