← sources

I tried different dataset sizes on KREA2, to try to answer the age-old question, how many images do you need?

source: reddit (reddit:StableDiffusion)published 2026-10-26 14:08ingested 2026-10-27 02:38

retained text

I created a synthetic dataset (She's an AI-generated woman not a real person.) and used those to train three different LoKRs for KREA2: one used 3 images only, the other 8, and one used 15 images. I used no trigger word in the caption for the training. I only addressed her as a young korean woman and I described everything around her in detail: her hairstyle, the background, the clothing, her facial expression, etc. 



The 3-image LoKR came out with a good likeness at 1,000 steps. The 8-image one at 1,440 steps and the 15-image one at 2,100 steps.



The workflow for creating the showcase shots: krea2 int8 with the qwen3vl 4b text encoder, refusal LoRA 1.0, turbo 0.4 (0.2 for styles), 16 steps, cfg 2, euler/beta57, shift 3, 2:3 at 2MP, Wan2.1 upscale2x VAE. For styles I usually increase the LoKr strength to 1.2 or 1.4 or 1.6.



The main takeaway (TLDR): for KREA2 you can make a LoKR that achieves good likeness with only three pictures. For retaining her likeness in different styles, you may need more pictures in the dataset. I could get her into some different styles with the 15-image LoKR but she doesn't have a very distinct feature that would distinguish her from being just a pretty woman with brown hair. (unless you count her stud earrings that show up in almost every picture, even style ones) 



The 15-picture LoKR wins at hard angles, shots from above, looking back over the shoulder, profile shots, etc 



But at most angles, like three-quarters, lower angle, beach shots, etc. All LoKRs tie.



I hope this is useful for anyone that had questions about this. As an example I always hunted and hunted for more and more high-quality shots of the character I wanted to make a LoRA of. At some point all you are doing is just slowing down your training and, if you are renting a GPU, then increasing the price it costs. 



Full write-up and every picture that I made for comparison in the comments below. It links to my site, just for disclosure. It's free but has my referral links. 

document fetched 2026-10-27 00:32 · canonical https://reddit.com/r/StableDiffusion/comments/1xfxuyr/i_tried_different_dataset_sizes_on_krea2_to_try