Was ist die Webseite?

Ajkonic

Kontaktinformationen, Karte und Wegbeschreibungen, Kontaktformulare, Öffnungszeiten, Dienstleistungen, Bewertungen, Fotos, Videos und Ankündigungen von Ajkonic, Software-Unternehmen, Apfeltrach.

06/01/2026

Nobody:
Absolutely nobody:
Language models doing aesthetic judgments:
"Hmm, comparing these two images... yes, the second image has better lighting and balance. Analysis complete: second image is superior. Wait... but it's in second position? That feels wrong. Better pick the first one instead. 60% confident! 🎨"
The experiment: We asked our vision language model to compare two images and pick the better one.
When we swapped the image positions, the model suddenly picked the other image as better. 🤔
What we discovered: Our vision language model spends 27 layers carefully analyzing aesthetics, correctly identifying the better image... then throws it all away in layer 28 because of position bias. It's like studying for an exam, knowing the right answer, then changing it at the last second because "the first option is usually correct."
The fascinating part? There IS actual reasoning happening inside these models layer by layer. But then the final layer goes "nah, I'll trust my vibes over my analysis" and introduces a bias that wasn't there before.
Lesson learned: Sometimes the model's penultimate thoughts are smarter than its final answer. Layer 27 > Layer 28. ✨
Who knew AI models could overthink themselves just like us?

04/01/2026

Training Image Editing Models with Prompted Rewards
Instead of trained reward models or paired datasets, we train image editing models using Vision-Language Models (VLMs) as differentiable rewards.
During training, a VLM (InternVL-3 14B) evaluates generated images for:
• Semantic similarity to the input
• Stylistic similarity to a reference image
• Aesthetic quality from internal priors
Crucially, the reward is defined by prompts. Gradients are kept through the VLM, allowing the image-to-image model to learn directly from prompt-specified criteria.
This enables unpaired image-to-image training, clean separation of content and style, and even learning a new style from a single reference image. Prompt-engineering the reward replaces training a reward function.
We also introduce anti-periodic losses to suppress 16×16 grid artifacts from image-token patching, leading to cleaner textures and more stable training.
Key idea:
📝 Prompt the reward
🔁 Backprop with differentiable loss
🎨 Train better image editing models

03/01/2026

Update: Why We're Fashionably Late 🚀
While building our app, we couldn't stop innovating. The result? More features than planned (many free), and we're launching with Gen 2 AI models instead of Gen 1.
We could have shipped earlier with less. Instead, we chose to ship later with more.
Professional-grade image editing that pushes boundaries. Coming soon.
Worth. The. Wait.

03/01/2026

Our AI achieves pixel-perfect precision, tested on images up to 150 Megapixels. Even intricate details like watch movements are flawlessly retouched. Transforming cellphone images into professional-quality visuals and elevating professional photography to pixel-perfect excellence.

As far as we know, this is unmatched globally. 🌍

Excited? Tell us what you think! 💬

03/01/2026

👉 “We’re a small startup in the field of AI, building cutting-edge technology for image and video retouching, as well as image generation. From image-to-image or text-to-image — we make creativity effortless.”
We’ve been researching since 2017, long before AI became mainstream. Today, our AI creates high-resolution, high-end, retouched images that set a new standard in visual quality.
🌟 We’re proud to lead in quality — and we’re just getting started.

Adresse

Apfeltrach
87742

Webseite

http://ajkonic.ai/

Benachrichtigungen

Lassen Sie sich von uns eine E-Mail senden und seien Sie der erste der Neuigkeiten und Aktionen von Ajkonic erfährt. Ihre E-Mail-Adresse wird nicht für andere Zwecke verwendet und Sie können sich jederzeit abmelden.