Sign In
Civitai turns 4. Pick up a hat.
homemodelsimagesvideos3D Modelshubsarticlescomicscruciblesupdatesshop

Soul Realism (Krea 2) - everyday phone photo LoRA

Updated: Nov 1, 2026

stylekrea 2amateurphone photophotorealisticrealism

Download

1 variant available

bf16 SafeTensor

soul_realism_krea2_v1.safetensors

BF16, good balance • 223.83 MB

Verified:

Type
LoRA
Stats

309

Reviews

(21)

Published

Nov 1, 2026

Base Model

Krea 2

Training
Epochs: 10
Usage Tips
Strength: 1
Hash
AutoV203316C214F

Krea 2 is licensed under the Krea 2 Community License Agreement. For more information, visit https://krea.ai/krea-2-licensing.

01_cover_square.png

Please read first

I made this for fun. There is nothing to sell you, and I'm not saying it's the best LoRA out there. Feel free to try it, and feel free to tell me what doesn't work. Below are a few tips that helped me get slightly better pictures.

It can make people look younger than you ask for. Always write a clear adult age in your prompt, and check your pictures.

What it is

A Krea 2 text-to-image LoRA that nudges pictures towards everyday phone photos: ordinary rooms, normal light, a bit of mess. Results vary a lot by prompt and seed. No trigger word.

Tips that helped me

  • Krea 2 Turbo, 8 steps, CFG 1, euler / simple.
  • Strength 0.8 to 1.0. Lower it if the picture gets too sharp or crunchy.
  • Describe the photo the way it was taken: "front-camera selfie", "photo taken by a friend", "mirror selfie", "flash".
  • Write the age, e.g. "a 28-year-old woman".
  • It likes adding tattoos. Almost half the training photos had them.

What it was trained on

2,136 unfiltered phone photos from around 110 people. No AI images, no beauty filters, no studio shoots, no memes. Near-duplicates removed. 10 epochs, rank 32, 512px, Fizgig trainer on Krea 2.

How I tested it

I tried some ideas from Krea's Krea 2 technical report. I made 27 versions (different epochs, strengths, and groups of blocks turned off or doubled) and rendered the same 8 prompts with 2 seeds each, 432 test pictures. GPT-6 Astra then looked at every picture on its own and answered simple yes/no questions: does it look like a real photo, real skin, real background, are hands and text OK, does it follow the prompt. This version (epoch 10 with blocks 21 to 27 doubled) got the most "looks real" answers. That is one AI's opinion on a small test, not proof. Plenty of pictures still look AI.

Known problems

  • People can look younger than prompted.
  • Some crunch / over-sharp detail.
  • Mirrors and reflections can come out wrong.
  • Only tested for text to image. Not tested with character LoRAs, so it may change how your character looks.

I'll keep working on it. Thanks for trying it.