Orbit · Explore

AI Lab · by Narin Chai · 2026-10-20

Image generation quality leaps: are we past the uncanny valley?

openai.com

Hands, text and lighting are mostly solved now. What still gives away generated images to you?

96 sparks · 21 comments

Comments

Ren Suzuki · 11 likes

Prompt sensitivity is still the biggest pain in production

Quiet Noodle · 0 likes

Yeah same here

Chinedu Adeyemi · 8 likes

Context window claims are always optimistic, test it yourself

Cian Kelly · 1 likes

Haha my thoughts exactly

João Pereira · 7 likes

Open weights or it didnt happen

Cedar Mango · 13 likes

Well said

Daniel Kiptoo · 11 likes

Depends but yeah mostly

Budi Santoso · 6 likes

Running a quantized version locally and its surprisingly usable

Cian Kelly · 8 likes

Good take, hadnt considered that

Brave Runner · 0 likes

Well said

Diego Quispe · 5 likes

Benchmarks look great but real world usage is where it falls apart for me

Mango Coder · 1 likes

lol this made me laugh

Noah Meier · 0 likes

Well said

Budi Santoso · 4 likes

Evals are getting saturated, we need harder ones

Léa Bernard · 12 likes

Well said

Tola Ibrahim · 2 likes

Not sure I agree but ok

Noah Miller · 2 likes

My GPU just cried reading this

Noé Petit · 9 likes

I read that twice and still chuckled

Paolo Garcia · 0 likes

The cost per token matters more than the leaderboard imo

Diego Quispe · 11 likes

Good take, hadnt considered that

Niamh Byrne · 4 likes

Not sure I agree but ok