Hacker News Reader

This is super cool. Using LLMs to mix together two songs that don't belong mixed together is a great usecase.

geooff_on Show HN: Turn raw HTML into production-ready images for free

Very cool. Is there an option to self-host? This seems like it could be a cool agent skill.

geooff_on FLUX.2: Frontier Visual Intelligence

Their published benchmarks leave a lot to be desired. I would be interested in seeing their multi-image performance vs. Nano Banana. I just finished up benchmarking Image Editing models and while Nano Banana is the clear winner for one-shot editing its not great at few-shot.

geooff_on Meta Segment Anything Model 3

Seems like theres no API access. Has anyone got the weights? I'm not sure what to fill out for `affiliation`

geooff_on I think nobody wants AI in Firefox, Mozilla

I'm looking for AI features in a few places, one example is a git client that can draft commit Summary and Descriptions. This should be a very simple task and could use a simple on-device model. It feels like AI features are a firehose though. You either get none and a very 2018 product experience or a complete rework of everything "designed for agents".

geooff_on Show HN: Geofenced chat communities anyone can create

After failing to create a hub (too large in my cade) the state of hub selection isn't reset (you need to cancel manually before retrying)

geooff_on Show HN: Springus – Outfit recommendations from your real wardrobe using AI

Extracted features from a vision model. I haven't experimented with CLIP yet but would like to as I think adding clothing search would be interesting

geooff_on Show HN: Springus – Outfit recommendations from your real wardrobe using AI

The app can handle everyday shots, as you'd expect though, poor inputs produce poor outputs. Theres really two components to this question though:

1. Can the app differentiate one article of clothing from background / other articles 2. Can the app group together identical articles of clothing

To answer 1. The app has decent performance with test set pixel level mean accuracy of 0.80 and mIoU of 0.69, the test set is all real world fit pix from myself and friends. The 0.8 is a bit misleading though as the errors often occur at clothing boundaries so in poor lighting there can be some border gore.

As for 2. this remains to be seen. Currently clothing aggregation (Grouping together two segmentations of the same shirt) is manual. I'm doing some studies on tuning cosign-sim thresholds but I think long term there may need to be a more robust approach.