Self-hosted, Self-Improving, Federated Everything App — dev management, agent orchestration, persistent mind, digital twin, second brain, and creative pipeline from a single dashboard over Tailscale to port the user into the OS
Warm video render batches
In Video Gen, Advanced → Renders in batch selects 1–20 separate videos using one prompt, model and set of conditioning inputs. Generate and Add to queue show the requested count. The default is one render.
- Leave Seed blank, or click its dice button, for an independent random seed on every render.
- Enter a seed for consecutive seeds: a batch of three starting at
0renders with0,1, and2. Seeds must remain within the unsigned 32-bit range; an overflowing batch is rejected before it is queued. - A batch owns one local queue slot until it finishes or is canceled. Each finished video is saved to history with its actual seed. Canceling or failing a later render preserves the earlier finished videos.
Warm batching currently supports the MiniMax H3 MLX runtime. One Python process loads its pipeline and adapters once, retains its model weights and modulation cache, and reuses the request's prompt embeddings in memory. Each video is rendered and saved before the next starts; decoded outputs are released between renders. This reuse does not depend on the optional disk prompt cache.
Other runtimes, cloud providers and federated targets do not offer this control. In particular, LTX's two-stage pipeline fuses an adapter into its transformer and unloads components between stages; looping that pipeline with retained weights would change subsequent renders. Support requires a runtime-specific contract, not merely launching the existing command repeatedly.
Batches produce independent clips, so they cannot be combined with chained
chunks or a scene-delivery tag. Batch records carry additive batchId,
batchIndex (zero-based) and batchSize metadata. Their durations do not train
single-render ETA estimates because startup is shared across the batch.