T2V
T2V is a FlashDreams v2 text-to-video application. It accepts an initial text prompt, keeps the selected model resident, and can start fresh generation sessions from an interactive prompt field.
Quick Start:
Controls
An interactive WebRTC run shows a prompt field over the latest generated
frame. Submitting it starts a fresh session and rollout cache without unloading
the model. Reaching --total-blocks leaves the final frame and prompt UI
active. Model adapters may add startup inputs such as a first-frame image.
Demo Arguments
| Argument | Description |
|---|---|
--prompt TEXT |
Set the optional initial prompt; explicitly empty text is rejected. |
--total-blocks N |
Set the number of autoregressive blocks; the selected model supplies the default. |
--device DEVICE |
Select the model device. Default: cuda. |
--compile, --no-compile |
Override model compilation. |
--seed N |
Override the diffusion seed for reproducible generation. |
--ui, --no-ui |
Enable or disable the interactive prompt overlay. |
Bidirectional models generate a complete clip in one block and require
--total-blocks 1. Image-conditioned implementations may additionally require
--image-path.
For example, write a generated clip to MP4:
uv run flashdreams-run-v2 DEMO_MODEL_SLUG \
--output-path clip.mp4 --timeout unbound -- \
--prompt "A cat surfing"
To keep the model resident and submit prompts from a browser:
uv run flashdreams-run-v2 DEMO_MODEL_SLUG \
--mode webrtc --host 0.0.0.0 --port 8089
Tests
uv run --no-sync pytest apps/t2v/tests -m ci_cpu