Misha Feinstein retweeted
Fibo 2.0 ranks #1 amonst open-weight model in our T2I benchmark including 7M human judgements (700K on open-weight matchups).
1.5K prompts, all outputs, matchups and votes can be inspected on Benchmark.ai
#Fibo2 is out today.
It’s the #1 open-weights image model on @ArtificialAnlys, and the weights are yours.
A few things worth knowing
🧵1/7
This took the whole company.
Research led by @MokadyRon pushed the model, engineering with @ItayAmir8 got it into production in days.
Thank you, team.
🧵6/7
Try it:
Bria Create: create.bria.ai/home
@fal : fal.ai/models/bria/fibo-2/ed…
@fal : fal.ai/models/bria/fibo-2/ge…
@bria_ai_ API: docs.bria.ai/image-editing
🧵end
Bria Expand.
Every image has a frame. This one breaks it.
Give it a crop, get the whole scene back. Original untouched, no prompt. Go tall or go wide.
In blind tests against the leading models, it won both ways.
Last week: every background model guesses, including ours.
So now you just tell it what to keep.
“only the helmet”
“the person and the mat, without the plant”
Guided Remove Background, by @bria_ai_ on @fal
What should I throw at it next?
Try it: fal.ai/models/bria/guided-re…
The full argument: medium.com/@mishaf55/there-i…
"Background removal is a solved problem."
Even I claimed it, with our own RMBG-2.0.
I was wrong. Here's why.
1/5
Most models guess the salient subject. Some extract one named object (like our ObjectExtract API)
Real intent is rarely one noun: "keep the woman and the mat, drop the plant, keep the shadow."
Nobody does that reliably today.
4/5
Background removal should take instructions. We're building that at Bria, releasing soon.
The full argument: medium.com/@mishaf55/there-i…
Send me the image your background tool always gets wrong.
