LightX2V Prompt Studio

Long video from a prompt: a 4-step DMD student denoises 3 latent frames at a time while the MemRoPE history cache keeps the scene coherent — 81 frames for a short clip, 1005 frames for the 60 s rollout. Frames stream in as they decode. In this browser the real Wan2.1-T2V-1.3B weights run on your own GPU (one transformer block per download, cached after the first load); browser prompts come from the gallery — arbitrary prompts need the local server, which encodes them from a gallery embeds file until UMT5 lands.

Prompt

engine: not connected

MemRoPE long-video rollout

No rollout running.

Output

0 / 0 frames

Frames play as they arrive — the decoder streams them chunk by chunk, so a 60 s rollout starts showing picture long before the last chunk denoises. In synthetic mode the model is random weights: the output is structured noise, which is the point — it proves the whole path (prompt → denoise → decode → play) without a download.