About This LoRA

This is a quick test LoRA for Krea 2. I trained it to see how model perform with big dataset with many different concepts.

About the Dataset

For this test I used 4000 real life photos from a bigger set of over 8000 images. I had used the full set before to train a model I call the Qwen Blackhole model.

You can find that model on Civitai named Qwen Overcooked here:

https://civarchive.com/models/2557806/qwen-overcooked?modelVersionId=2874469

If you feel like comparing the two, go ahead and check it. I will add more example pictures for this LoRA when I get some free time.

Dataset images come from DSLR cameras, mirrorless cameras, phones, and other digital cameras.

Usage Tips

This is an early test and is not fully polished yet.

If you mix it with other character LoRAs, try lowering the strength to around 0.3–0.5. That often gives better results.

Training Details

Dataset: 4000 images

Resolution: 1024

Network Rank: 256

Batch Size: 8

Epochs: 8

Note

I have been training LoRAs and fine-tuning models as a hobby since the days of SD 1.5. My last big project was Qwen 2512, and it took months to gather the images, organize them, prepare everything, and train.

I recently noticed Krea 2 and it looked interesting, so I ran this quick test. Instead of keeping the result to myself, I put it here so others in the community can try it before I expand the dataset and do a proper full finetune.

I work six days a week, so I dont have a lot of free time for my PC. But I still experiment when I can.

I train and fine-tune models as a personal hobby and to keep learning. I share what I make just in case its useful for someone else. I believe in open source, trying things out, and sharing what you learn.

This LoRA is an experiment, not a finished product. You are welcome to try it, compare it, or just skip it if its not what you need. Either way, thanks to anyone who takes the time to test it and share honest feedback.