The LongBanana benchmark

Case Gallery

Explore the references. Read the instruction.
See what it takes to get every detail right.

Explore 90 cases View leaderboard
90 cases9 categories32 references / case

Reading a case

One instruction.
A complete visual context.

Each case includes its full synthesis instruction, a simulated target, and 32 reference images. Targets illustrate the intended scene; they are not a model leaderboard result.

Learn how outputs are evaluated