
What happened
An independent developer published auroray, a roughly 200,000-parameter MLP filter that draws candidate object-boundary lines on images. It is trained on zero human labels, using only programmatically generated geometry data, and is positioned as an AI-based alternative to OpenCV.
Why it matters
Because the filter only draws boundary candidates and knows no object categories, it can catch boundary candidates even on objects it has never seen, according to the developer.
Summaries like this, every morning.
auroray sits deliberately against the current grain of very large general-purpose models. The developer's stated underlying question is how small a large network can be made, citing concerns around power consumption, compute cost, difficulty of transferring between models, and dependence on giant infrastructure and very large companies. A tiny, function-specific filter, the argument goes, is easier to embed in existing systems and easier to tune or retrain where large compute is not available.
The core V1Filter takes a 9×9×3 HSV patch and outputs a single boundary score for the center pixel. Training data is generated rather than labeled by hand: clean boundary patterns are created first and then degraded — lines broken or recolored, noise mixed in, patches partially erased — to build robustness. Negative cases explicitly teach the filter not to fire on boundaries that miss the center or do not connect edge to edge. The developer notes the filter can also pick up lines that are not object boundaries, such as wrinkles or patterns on clothing, and says strengthening boundary-candidate features was prioritized first.
A comparison tool pits V1Filter against OpenCV output, and a GUI tool overlays extracted boundary lines on a captured region in real time. Looking ahead, the developer wants to extend the same lightweight approach to overlap detection and centerline extraction, and eventually to replacing parts of CNN layers and building a general image classification system that pairs small filters with a vector database for easier additional training.
Pick your industry and the AI tools you use, and get news related to your work every day.
Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →
Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.
Jane Street published socratic-tutor, a 102-line text file in its agent_skills repo, on 2026年11月3日; a test on…

Fixstars began offering GLM-5.3 on Fixstars Inference Cloud on November 20, using an OpenAI-compatible API; th…

On Anthropic's IPO day, Fireworks AI co-founder Chenyu Zhao told TechCrunch that open-weight models trail fron…

Nvidia is in advanced talks to buy Inception, a two-year-old diffusion-based language-model startup, for more…

Paris-based Remanence raised a $30 million seed round co-led by EQT Ventures and FirstMark, with Zero Prime, E…

Resolve AI announced SRE-1, a 27 billion-parameter open-weights model trained by 16 researchers over 7.4 days…