Krea 2 Realism LoRA (by gokaygokay)

This is not my LoRA.

I am republishing Krea 2 Realism LoRA by Gökay Aydoğan, so it is easier to find and use on CivitAI. All credit for training, design and the original release belongs to him.

Original repository: https://huggingface.co/gokaygokay/Krea-2-Realism-LoRA

Creator profiles:

- Hugging Face

- CivitAI

- X: @gokayfem

Weights file on Hugging Face: krea2_realism_lora.safetensors

License on the original card: krea-2-community-license

Trigger Word: none (plain natural language)

Recommended Weight: 1.0 (works well from 0.7 to 1.0)

The LoRA pushes Krea 2 Turbo toward candid, true-to-life photography: everyday people, lived-in rooms, motivated light, and the small mess of real life. You do not need photorealistic, 8k, DSLR or similar style tags. A longer descriptive sentence is enough.

About this version

This is a community rehost of the original Hugging Face release by Gökay Aydoğan. It was trained with the FAL Krea LoRA Trainer on krea/Krea-2-Raw and is meant to be used directly on krea/Krea-2-Turbo (Krea’s recommended Raw-train / Turbo-run workflow).

No trigger word. No style keyword diet. Just describe the scene.

What it does

Krea 2 Realism LoRA steers Krea 2 away from generic “pretty AI photo” and toward documentary snapshots: steam on a bathroom mirror, grease on overalls, mismatched mugs, laundry on a balcony line, a work lamp over an open engine.

Expect:

- natural skin and imperfect faces

- believable indoor and outdoor light (window, lamp, dusk sky, haze)

- clutter that feels used rather than staged

- people doing something specific instead of posing for a catalog

It is strongest on ordinary life: work, home, streets, shops, breaks, weather.

Trigger Word

There is no trigger word.

Do not add in the style of …, photorealistic, shot on iPhone, masterpiece or 8k. Those usually make the result stiffer. Write one flowing sentence: who is doing what, where, under which light, plus a few tactile details.

Recommended LoRA Strength

0.7 – 0.8: Subtle documentary pull. Use this if faces get too harsh, rooms too messy, or you are stacking other LoRAs.

1.0 (recommended): Full look from the original card. This is the strength used for the official Hugging Face examples.

1.1 – 1.2: Only if the base checkpoint stays too “illustrated”. Can oversharpen pores and dirt. Start at 1.0.

Recommended generation settings

These come from the original model card:

- Base: Krea 2 Turbo

- Steps: 8 (Turbo)

- Guidance / CFG: 0.0

- LoRA scale: 1.0 (or 0.7–0.8 for a lighter touch)

- Resolution: short side 1024–1536; official examples are 1280×1600

Portrait / 4:5 / 3:4 matches the training examples best.

How to prompt

Best results come from a single long sentence:

[person] [specific action] [place], [motivated light], [2–5 lived-in details]

Good details: water droplets, worn wood, grease stains, laundry, ticket stubs, steam, dust, mismatched objects, a lamp that actually belongs in the room.

Avoid: tag soup, multiple competing lights, empty studio backdrops, and “looking at the camera” unless the scene is really a selfie.

Example prompts

A young woman taking a selfie in a dim bathroom after a night out, phone covering part of her face, only the vanity bulbs lighting the steamed mirror and a damp towel slipping off her shoulder

A barista with tattooed hands wiping the steam wand of an espresso machine in a tiny cafe just after closing, the last warm light from the street hitting rows of unwashed cups on the back shelf

Model Compatibility

Optimized for Krea 2 Turbo. That is the checkpoint named on the original card.

It can be stacked with other Krea 2 LoRAs, but drop this one to 0.7–0.8 if a second style LoRA is already doing heavy lifting. It is not a Z-Image / Flux / SDXL LoRA.

Sample gallery notes

Gallery images on this page were generated with this LoRA on Krea 2 Turbo at strength 1.0, following the prompting style above. They are demonstrations of the original weights, not a claim that I trained the adapter.

Credits

Author / trainer: Gökay Aydoğan ([gokaygokay](https://huggingface.co/gokaygokay))

Original release: https://huggingface.co/gokaygokay/Krea-2-Realism-LoRA

Trained with the FAL Krea LoRA Trainer on krea/Krea-2-Raw, for inference on krea/Krea-2-Turbo.

Base model credit belongs to Krea. LoRA credit belongs to Gökay Aydoğan.

This CivitAI page is a convenience mirror and usage guide only. I did not train these weights and I am not claiming authorship. If the original author publishes an official CivitAI version, use that and treat this page as a pointer.