Today we're introducing SimForge Factory. Pick SimForge scenarios and the looks you want, such as night city lights or a wet road after rain, and Factory turns each scenario into realistic camera video, checked against the scenario and labeled from it.

The events that matter most for a driving model are the ones a fleet rarely records: a cyclist cutting in, a car coming the wrong way, someone crossing in the dark. In SimForge you script an event once, on a real street. Factory then shows it in many kinds of light and weather, and every clip keeps the scenario's ground truth, so its labels come from where things actually are, not from someone drawing boxes.

One scenario, eight looks.

Choose a scenario. The centre tile is its SimForge render; around it are the eight clips Factory generated from that scenario, one per look, playing in step. They are shown as delivered, including where the generator got things wrong.

  1. Clear noon
  2. Golden hourExperimental
  3. Low winter sun
  4. Overcast
  5. SimForge render
  6. Wet after rain
  7. Heavy rainExperimental
  8. Rain at duskExperimental
  9. Night city lights

The person crosses in every clip, and the check accepted all eight. Some clips also add cars or people that are not in the scenario.

On San Francisco · Mission, a private World not in the open release. Each clip is 20 s from the front wide camera; every generated clip is billed, flagged ones too.

How it works.

  1. Render. SimForge simulates the scenario on its World and renders it from each of the four cameras.
  2. Generate. For each camera, a layout video shows where every road user is. Factory runs NVIDIA's Cosmos3-Nano-Sim-Auto video model on our GPUs to generate a realistic clip that follows that layout, in the look you chose.
  3. Check. Each clip is compared with the render: the scenario's key road users must still be where the scenario puts them, and the look must have applied. The check labels each clip accepted, flagged or rejected, so you can see what the model got wrong.
  4. Label. Labels come from the scenario, not from the picture: a box for every road user in every camera, their positions in 3D, and the path of our own car.

Generated video is not perfect. Some looks are experimental and may not apply well, the model can add road users the scenario does not have, and the check verifies the key road users and the look, not overall image quality.

What you get.

  • Up to four cameras. A 20-second clip from the cameras you choose of a four-camera rig, generated together, so the views agree with each other.
  • Labels from the scenario. Per-frame 2D boxes for each camera, 3D boxes for every road user and our car's path, as Parquet files.
  • What went in. The layout videos and frames the model was given, so you can see what each clip followed.
  • The check's result. What was checked in each clip, and why it was accepted, flagged or rejected.

One clip, 4 cameras. The city crossing in Night city lights, every camera of the rig generated together and playing in step.

The same scenario, also as simulated LiDAR from SimForge. Factory makes camera video only; SimForge's own renderer renders the scenario's LiDAR on your machine with the CLI. Sensor examples

Choose your cameras.

A scenario clip uses the standard 4-camera rig, and you choose which of its cameras you need, from one to all 4. You pay per camera: at 4¢ per camera-second, each camera of a 20-second clip is $0.80, so a clip from the front wide camera alone is $0.80 and all 4 come to $3.20.

  • Left side, not chosen
  • Front wide, chosen
  • Right side, not chosen
  • Front telephoto, not chosen

In the CLI, name each camera with --camera:

simforge factory clip create --scenario SCENARIO_ID \
  --look night-city --camera camera_front_wide_120fov

A raw generation can use any of the model's 11 camera views, in the order you choose, each with your own layout video. You can render those layouts from your own SimForge rig, with each camera mapped to one of the model's views. The views themselves are fixed: the model knows only the 11 it was trained on.

  • Front wide
  • Right side
  • Rear right
  • Rear telephoto
  • Rear left
  • Left side
  • Front telephoto
  • Front fisheye
  • Left fisheye
  • Right fisheye
  • Rear fisheye

Availability and pricing.

SimForge Factory is available on SimForge Cloud. SimForge Cloud is open in beta. A clip costs 4¢ per camera-second: a 4-camera, 20-second clip is $3.20. Every generated clip is billed. Our automatic check labels each clip accepted, flagged or rejected, so you can see what the model got wrong. A clip we fail to make is free. A raw generation, where you bring the model's inputs yourself, is $1.00 whatever it shows, with no check. New accounts start with $25 in free credits, and you can buy more on your workspace's credits page.

Get started.

Sign in at simforge.ai and open Factory in the Studio, or run it from the SimForge CLI:

simforge factory looks
simforge factory clip create --dataset DATASET_ID \
  --look night-city --look wet-after-rain --dry-run

--dry-run prices the run without starting it. Without it, the CLI shows the price and asks before it starts.

Runs on SimForge Cloud.