Tired of the plastic, airbrushed, "perfect AI" look? This LoRA brings back the unfiltered look of a real phone snapshot: harsh flash, bad white balance, soft focus, blown-out windows and awkward framing. Nothing is polished, and that's the point.
Trained on naturally captioned amateur snapshot datasets for Krea 2 and Qwen-Image-2.1.
What it does
- Real camera flaws: low-light noise, soft focus, lens smear, flare, blown highlights and uneven auto white balance.
- Real lighting: direct flash, mixed room light, streetlamp color casts, and backlit window bloom.
- Unposed compositions: casual mirror selfies, candid street shots and awkward crops, with no studio feel.
Versions
Qwen-Image-2.1 (v1)
Qwen works best with natural-language paragraphs. Describe the scene and light casually.
- candid_snap photo of a young woman wearing a casual grey hoodie sitting in a dim diner booth, direct flash lighting, holding a coffee cup
- candid_snap amateur snapshot of a man walking down a rain-slicked city sidewalk at night, yellow streetlamp reflections, harsh ambient lighting
Krea 2 (v1.0)
Use descriptive tags alongside the trigger word.
- r3alism, a casual flash mirror selfie of a woman in her bedroom, messy hair, low-light phone camera texture...
- r3alism, an amateur smartphone snapshot of a person walking on a city street at night, visible digital noise...
Krea 2 (v2.0)
V2 isn't a replacement for v1. It's a different take. V1 gives a stronger composition and style out of the box, and V2 leans into a softer, blurrier amateur look with closer prompt adherence.
V2 was retrained on the same image set with rewritten, more descriptive captions and a lower rank.
- The look scales with your prompt: ask for flash, grain or a blown-out window and it delivers. It doesn't force a night color grade onto daytime scenes.
Krea 2 (v3.0): WIP
In progress and not released yet. Goals:
- Sharper, more readable faces.
- More realism in general, including daylight photos.
- Wider variety: street scenes, food, car interiors and everyday objects, not only selfies.
- Better behavior alongside character LoRAs.
Follow along for updates. These are goals, and I'll update this page once V3 is tested.
Description
FAQ
Comments (13)
Making some rendes and the LoRA is performing very well, thanks :)
Thank you 🙏 I’m glad you like it!
How many images did you train this on?
82-images, captioned with Qwen-3-VL, trained on Krea 2 Raw
@kotsarelos62662 Okay thank you, where did you find the images?
@MonkeyForever Pinterest
@kotsarelos62662 I see, one last question, how did you caption it, like did you just use the default captioning prompt in ai toolkit?
@MonkeyForever Nah, I didn't use the default prompt. I used a custom prompt specifically written to capture camera traits like direct flash, sensor noise, and lens glare in a clean 120-token paragraph. (Trained it over on TensorArt, but I guess the captioning concept is the same).
@kotsarelos62662 thats interesting cuz people say you shouldn't include camera stuff and "style" stuff in the caption when doing a style lora.
@MonkeyForever Honestly, I debated that exact same thing before running it.
My reasoning was just versatility. The way I see it, you tag stuff that you want tied to specific words, and leave untagged whatever you want baked in as the baseline style.
If you leave camera traits untagged, it bakes grain and harsh flash into every single picture no matter what. By tagging them, people can prompt for stuff like direct flash or sensor noise when they want that heavy look, or leave those words out if they just want a cleaner casual shot.
Just want to let you know, I tested this on NSFW loras and my own character loras and they work perfectly. I didn't even have to lower the weight!
Thank you so much.
@Tezozomoctli Thanks for letting me know, appreciate it! I specifically trained it with versatility in mind so it wouldn't crush character details or mess up other LoRAs, so hearing that it runs clean even at full 1.0 weight means a lot. Enjoy!
UwU