Video

Prism's Maker/editor office covers generation and generative transformation of visual material through time. A video model may create a shot, animate a still, continue a clip, transfer motion, or condition performance on audio. It does not thereby own visual direction, sound production, the final audiovisual timeline, structured animation curves, or publication.

This candidate study was reviewed on 2026-08-26. It records a proposed contract, runtime shortlist, model register, and proving bake—not delivery, automatic fallback, or permission to generate, transform, export, or publish video.

One temporal job, several proved operations

The inner surface should remain stable while engines and models evolve. A candidate VideoJob@1 names one explicit operation:

Optional material never invents support. Each model profile declares the operations, source kinds, reference counts, controls, output limits, and cancellation behavior that its exact bake proved. First/last-frame control, masks, camera paths, poses, depth, edges, generated sound, interpolation, and upscaling are independent claims rather than implications of generic video support.

Sight may supply source-grounded masks, regions, pose, depth, flow, or tracks as exact controls. Video owns the generative temporal effect; a track remains a bounded estimate with gaps, not subject identity or proof of what happened.

Kinesis owns reusable skeletal, root, hand, face, contact, and transform motion facets. A motion-transfer or audio-driven video may visibly perform while returning pixels only; it does not imply a rig-compatible clip. If one worker returns both pixels and structured motion, the compound result retains separate Video and Kinesis facets, validation, and provenance.

The job carries immutable source and control ArtifactRef values; original and derived prompts; negative prompt; seed policy; width, height, frames, frame rate, duration, and timebase; an immutable preset; deadline and budget; output profile; and a closed engine-extension object. A prompt translator may derive the model-facing prompt from the operator's preferred language, but the receipt preserves both texts, translator revision, and declared loss. Translation never silently replaces visual direction.

VideoJob@1 owns the requested temporal effect, candidate set, technical validation, and result settlement. Creative adoption belongs to Voidlight or another consuming Composition; successful custody and probes never accept the work on that owner's behalf. Each concrete execution uses Core's Designed ServiceJobAttempt@1 mechanics. Numeric progress, step progress, previews, cancellation before start, cancellation while running, and cancellation granularity are separate Connector facts. Deleting a provider record is not evidence that a running generation was cancelled; an indeterminate attempt remains contained and is reconciled by the same provider identity.

Every returned file is immediately ingested into Reliquary custody and rehashed. A canonical receipt retains input and output digests; engine, container, model, encoder, decoder, VAE, LoRA, and control revisions and licenses; graph or preset digest; sampler, steps, guidance, seed, dtype, quantization, offload and parallelism; dimensions, frames, rate, timebase, codec, timing, warnings, cost, progress, and cancellation settlement. Provider IDs, filenames, and embedded metadata are evidence inputs, not canonical provenance.

Serving route, workflow route, and deterministic tools

Diffusers remains a model SDK and compatibility substrate, not another connector unless LychD deliberately builds and accepts a durable worker around it. OpenAI Videos is likewise a useful wire dialect, not the inner contract: LychD keeps job identity, authority, custody, retry, cancellation, and recovery.

A Comfy graph is an engine program, not a Spellweaver Pattern or another workflow jurisdiction. Its Rune selects an immutable LychD preset whose graph, nodes, model dependencies, parameter openings, network behavior, and output paths passed Assimilation. Runtime downloads, arbitrary imported workflows, ambient custom-node installation, and unapproved partner nodes fail closed.

FFmpeg is deterministic machinery, not a generative model or application. Prism's Lens may use a pinned FFmpeg revision for per-artifact probing, extraction, normalization, interpolation, or encoding with declared loss. Voidlight owns creative motion and accepted visual sequences; Broadcast owns the final audiovisual timeline, trim, placement, captions, mix, mux, and editorial render. A generative v2v operation never becomes the final cut merely because its input was already video.

First model profiles

The initial register keeps one permissive routine family and a few materially distinct specialists. These are exact profile candidates, not claims that one family name grants every operation.

Wan TI2V-5B is the first permissive routine default; VACE is the first controlled editor. Animate and S2V remain distinct specialists because a general T2V/I2V checkpoint does not prove motion transfer or audio-driven performance. LTX remains valuable where native joint sound and video outweigh its license and resource burden. Hunyuan is neither globally accepted nor globally banned: it is ineligible where its own territory clause withholds a license.

If a model emits sound with video, Prism records one compound candidate, original container digest, shared timebase, and generation receipt, then separates the returned streams into attributable child artifacts when their rights and formats permit it. Voidlight retains visual judgment; Riffmaw musical judgment; Language Edition timed-language judgment; and Broadcast picture-sound and final audiovisual judgment. Unexpected audio remains a typed unexpected facet or quarantined candidate. generate_sound=true cannot silently transfer any office to a video connector, and inseparable streams are accepted or refused as one compound use rather than laundered by demux.

Profiles, Runes, Covens, and arbitrary iron

Hardware suitability belongs to an exact deployment profile, never to the universal job contract. The same model may be permanently resident on a large-memory workstation, offloaded or quantized on a smaller host, sharded across measured devices, or reached through an explicitly admitted Portal. A personal two-card bake is evidence for that host, not a product ceiling.

The service path remains distinct:

VideoJob@1
→ CapabilityDemand(interface, operation, typed facets, eligible profile refs)
→ Dispatcher issues JobGrant or HardwareTransitionRequired
→ Orchestrator converges scarce local iron when required
→ re-dispatch and invoke the exact granted driver

A Designed Soulstone Rune declares the concrete service instance and exact [[capabilities]] references to immutable interface/profile/driver/dialect/evidence/resource definitions, plus devices, mounts, lifecycle, group membership, and conflicts. Current [[models]] capability hints belong only to v1 model compatibility. A Coven only names compatible Soulstones that may rise together. It does not choose placement, merge GPU memory, schedule jobs, relax conflicts, put another service to sleep, or authorize Portal fallback.

One deployment may keep a compact Mind and routine Wan worker on separate consumer GPUs. Another may keep a large Mind and a heavy video specialist together on one high-memory workstation. A third may drain incompatible leases, invoke runtime-native sleep through Orchestrator, and assign several measured devices to one worker. Two devices do not become pooled memory without exact runtime support, and multi-device execution is admitted only when its chosen tensor, pipeline, sequence, or stage topology proves correctness and benefit.

vLLM-Omni Sleep Mode can release most stage VRAM while retaining process state, making it a strong future runtime transition mechanism. The Run—not the engine—enters Stasis; Orchestrator owns lease drain, transition admission, readiness, restoration, and uncertainty containment.

The proving bake

Promotion is per exact engine, model, dependency, preset, precision, quantization, offload, parallel topology, device class, resolution, frame count, and rate. The matrix includes a bounded 24 GB consumer profile, a single large-memory workstation profile, an explicit multi-device profile, and a Portal profile where policy permits. No universal VRAM formula substitutes for receipts on operator iron.

The corpus covers multilingual and Slovak briefs with preserved prompt derivation; text-to-video; single- and multi-image conditioning; first/last frames; continuation; controlled and masked V2V; identity and character motion; speech, song, and music-driven performance; camera and temporal continuity; subtitles and visible text; adversarial dimensions, codecs, metadata, and durations; cancel before start and during generation; OOM; restart; stasis/wake; and deterministic-seed replay where claimed.

Measure prompt and control adherence, source and identity preservation, temporal coherence, motion quality, lip and audio synchronization, flicker, visible text, accessibility hazards, duration and timebase correctness, output validity, latency, peak VRAM, host RAM, disk traffic, warm-up and wake time, throughput, cancellation settlement, recovery, artifact lineage, and license closure. Promote only the exact profiles whose measured costs and failure behavior remain honest; more powerful iron may admit a broader resident Coven without changing VideoJob@1.