CUT!Best-of-N Studio
15-minute self-guided lesson
NeurIPS 2026 · Education TrackVisual primer · no code required

A visual lesson for autonomous driving

How do you test weather
you cannot schedule?

THE PROBLEM Critical cases are scarce

A real test fleet records whatever the road provides. It cannot safely or reliably stage dense fog, severe glare, or a nighttime hazard on demand.

EMERGING METHOD Build controlled counterfactuals

Teams use simulation and generative world models to expand coverage beyond recorded miles.

CONCRETE IMPLEMENTATION NVIDIA Cosmos Transfer 2.5

Transforms a recorded drive into controllable weather variants, using spatial guidance to preserve the scene across time.

Synthetic variants can probe a data pipeline. They do not certify vehicle safety.

01 · MISSING CASES

Rare weather is hard to schedule.

Rare combinations of weather, light, traffic, and place matter for validation, yet they are difficult to capture safely and repeatedly.

long-tail coverage
AV DATA COVERAGETHE LONG TAIL
WHY THIS MATTERSRare conditions cannot always be safely staged or reliably collected.

Demo outputs use Cosmos Transfer 2.5. NVIDIA directs new model development to Cosmos 3.

Start the audition
01 · DIRECTOR'S CHALLENGEAbout 8 seconds

The judge is missing

Which generated video should we keep?

The winner enters an AV training set. Pick the version with convincing fog and an intact street.

TAKE 03A
TAKE 07B

CAST YOUR VOTE

Which candidate would you keep?

02 · THE AUDITION

How many candidates can the judge inspect?

Increasing N expands the search. It also creates more work for the judge.

Sample budget, N1 take

N=1 offers no comparison. The first sample ships.

1 take on screen

YOU ARE THE JUDGE

Inspect the visible candidates. Select a take.

05 · CHANGE THE WEATHER

Roll it again: rain and snow.

Reuse the same 4 seeds. Change only the weather prompt. Variation remains.

“Moderate fog fills the street… distant buildings are softened.”

Compare takes under the same weather condition.
OPTIONAL · BEHIND THE CAMERAHow were the videos made?Open production notes +

Cosmos generated each candidate. Ray ran 4 GPU actors. 8 seeds finished in 2 waves.

NVIDIA COSMOSvideo generation + RAYparallel actors
@ray.remote(num_gpus=1)
class CosmosActor:
    def audition(self, seed):
        return cosmos.generate(seed=seed)

actors = [CosmosActor.remote() for _ in range(4)]
takes = ray.get([
    actors[i % 4].audition.remote(seed)
    for i, seed in enumerate(seeds)
])
winner = select(takes, rubric)
Source videoThe scene to preserve
Edge controlThe structure to preserve
Depth controlAlternative structural control

Transfer 2.5 supports several structured controls. This experiment used the distilled edge checkpoint. The depth view above is illustrative.

1Conditionsource + prompt + edge
2GenerateCosmos Transfer 2.5
3Parallelize4 Ray actors · L40S
4Selectrubric + evidence
8 / 8successful fog generations
106 smean generation time
474 sobserved wall clock, including loads
42.1 GBGPU memory per actor
Show the exact fixed settings
Model edge/distilledSteps 4Guidance 4Edge weight 0.65Frames 93Resolution 1280×720 generated

06 · CHECK THE CUT

Got it?

2 questions. 30 seconds.

QUESTION 1 / 2

What changed across the 8 fog takes?

QUESTION 2 / 2

Why might a larger N fail to improve the final output?

0 / 2

Complete the knowledge check.

THE TAKEAWAY

Generation creates options.
The judge turns options into a decision.

The generator-selector pattern now appears in reasoning, generation, alignment, and evaluation.

Run the audition again ↑

Sources & production notes

Built on NVIDIA Cosmos. Cosmos Transfer 2.5 source uses the Apache-2.0 license. Model weights use the NVIDIA Open Model License. These videos are educational outputs from the published example input. No endorsement is implied.

A FINAL LOOK

Rain. Fog. Night. Clear.