Action2V
Action2V is an interactive FlashDreams v2 application for world models driven by direct keyboard and mouse actions. A selected model supplies its pipeline, first-frame resolution, and action-vocabulary mapping.
Quick Start:
Controls
Action2V preserves held keyboard and mouse-button state. Pointer motion and wheel deltas are consumed once per model action. Reset and focus loss clear held state and pointer history.
The selected model maps those model-neutral inputs to its own action
vocabulary. Use --reset-key to choose the ASCII letter that resets the
current session.
Demo Arguments
| Argument | Description |
|---|---|
--image-path PATH |
Use an RGB or RGBA image to establish the initial world state. |
--example-data, --no-example-data |
Enable or disable the selected model's example image. |
--device DEVICE |
Select the model device. |
--total-blocks N |
Stop after this many generated action blocks. |
--ui, --no-ui |
Enable or disable interactive pointer capture. |
--seed N |
Override the model RNG seed for reproducible generation. |
--mouse-sensitivity SCALE |
Multiply pointer motion by a finite, non-negative scale. Default: 1.0. |
--reset-key CHAR |
Reset the current session when this ASCII letter is pressed. Default: T. |
Model-free demo
Use the packaged example image to exercise Action2V without loading a model:
uv run --package flashdreams-action2v flashdreams-run-v2 action2v-dummy \
--mode webrtc -- --example-data
Tests
uv run --no-sync pytest apps/action2v/tests -m ci_cpu