RelateAnything
Open-vocabulary relation prediction between any image regions, in real time
Emerged finds new products worth a look, daily — what’s new today
If you build search or RAG over screenshots, PDFs and video, this gives you one open embedding space for all of it instead of stitching separate models together.
WeMM-Embedding is a family of universal multimodal embedding models (2B, 4B and 9B) that map text, images, video, visual documents and interleaved inputs into a shared vector space. Weights are on Hugging Face under Apache-2.0, with Matryoshka support so you can trade dimension for storage. The authors report leading results on MMEB-v2 and MMEB-v3, where the 9B model scores 59.5 overall.
More like this
Jinfer
An AI inference engine that runs inside the JVM, with no Python sidecar
L
Laya Vision
A 256M model that answers typed questions about an image with calibrated odds
DigUp
Describe the thing you half-remember and it opens the right page
Overtone
Store each spoken word once, then let the player pick the line breaks
rea-video
An agent skill that works out how a video was made and how to remake it