This lora recreates the look of the Siemens CX70, an early-2000s 0.3 MP CCD camera phone. It focuses on authentic cx70 color rendering, sensor noise, highlight bloom, soft rendering, and vintage digital imperfections, bringing the nostalgic feel of pre-smartphone photography to any subject.
Trained on 13 images from https://vk.com/siemensphotography
The dataset was expanded to 32 images for v2.0.
There is a different trigger words for each version/model
Qwen2507
More of a conceptual than a physical representation of camera effects. A well-visible effect, y2k vibe, slightly sloppy style of qwen2507. 1mp works best.
Flux.2dev
Excellent physical representation of the Siemens CX70 camera, good noise distribution, color selection, white balance, composition, and a slightly better understanding of liminality and doomer aesthetics. 0.3-1mp works best, but it can do all flux's 4mp and apply effect on images.
Krea2
The dataset was expanded to 32 images.
Works best with:
1mp
1.2 strength (for v1.0)
Krea2 RAW + 0.6 turbo lora
in v2.0 trigger words was removed. In most cases, the model applies the effect even to other styles, but sometimes it needs to be forced using prompts like "40mm lens f/2.8, VGA CMOS sensor aesthetic, 0.3 megapixel resolution. Digital photography, significant image noise, heavy grain, slightly overexposed." This is an issue with Qwen-like models.
I will be very happy to see your ideas that came to mind when exploring the possibilities of this LoRa, so feel free to share your generations! :>
Description
FAQ
Comments (13)
really nice on krea2. At 1.0 you get the full experience, which looks like it'll be full of VAE artifacts in the preview, but ends up looking rich and consistent. My only real complaint is it can trigger inconsistently if I don't use the trigger words -- which can be confusing sometimes.
Hey, thanks for the feedback!
Qwen-like models don't train very well on this style with a single (unique) trigger word, but I plan to fix this in future versions.
If you make a update at some point, please consider shortening the trigger to something simple, if that's possible. I try to put trigger words/phrases into the filenames of loras but with such a long one I can't. Something unique like CCDrmz might work. I don't know if Krea 2 can be trained to use a trigger like that for the complete trigger stuff that you want or not, though.
Done
(almost)
looks good, i just wanted to release my nokia lora :))
btw, siemens cx70 has VGA CMOS Sensor, not ccd (⌐■_■)
Crap, I completely missed this :d.
At first, I made this LoRA just for myself, but it turned out so well that I decided to share it, and then I just got hooked into it)).
Right now, I'm exploring what I want—or at least what I can—train, so I don't have the time to change the dataset. But ultimately, my goal is to get it down to a single trigger word.
Thanks for the feedback! :>
Just made a bunch of stuff using ur lora, can i ask how you captioned it? How many steps? This is a remarkable lora
Hi, thanks for the feedback! I used Qwen 122B A10B IQ4_XS to make the captions. I did about 90 epochs with 32 steps each, 2880 steps total.
@Fenfel23 May I also ask, how detailed and long were the captions? were you focusing on the tech siemens lens look? I never tried a concept lora and yours looks even better than danrisi's
@meneper This is what i meant
@Fenfel23 Okay thank you, where do you train, cuz im abit confused about the epochs part and "steps each", all i can do in toolkit is just set the total steps.
@Fenfel23 can you please show me your captions i really need to know how to caption properly
@Fenfel23 can you please show me your captions you used to train on the model? i really need to know how to caption loras properly, thank you!