4.6 KiB
RFC: Studio as the creative suite — every content type, one flywheel
- Start Date: 2026-07-18
- Author(s): @zooqueen
- Status: Draft
- Related: studio-agentic-creative-runtime.md
Summary
Studio today ships one content lane — image — end to end: generate from CAD or reference, organize in Stacks, fix with lineage, rate with feedback that conditions the next render, quality-gate every output, render on your own GPU or ours. This RFC makes that lane the TEMPLATE for every creative content type — video, music, decks, documents — so Studio becomes the suite: Canva's breadth with a generation flywheel Canva does not have. One library, one queue, one feedback loop, one wallet; a lane per medium.
Motivation
Teams do not want five tools with five libraries. The differentiators Studio already has — provenance per job (graph, seed, params), the rating→conditioning flywheel, bring-your-own GPU — apply to every medium unchanged. The cost of a new medium is a lane, not a product.
Design
The lane contract
A content type is ONE lane implementing four seams; everything else is shared.
| Seam | What it is | Image lane (shipped) |
|---|---|---|
| generate | graph templates the composer + templates grid dispatch | qwen-image-edit family |
| card | how the library renders it (thumb, player, badges) | .card thumb |
| edit | the surface "Fix"/"Open in Editor" opens | fix dialog + node editor |
| judge | what the quality gate asks the vision/audio model | photoreal skin/design gate |
Shared, untouched by new lanes: library + Stacks + Archive, worklog identity (unique per-job prefixes), queue + history + lineage, ratings → do-not-repeat conditioning + feedback.jsonl, wallet + GPU fleet, org tenancy, analytics.
Lane order and reality
- Video — ✅ SHIPPED (2026-07-18). The Wan 2.2 TI2V-5B lane over the ONE
library/queue/flywheel:
_video_graph(proven node-for-node on spark, ~154s for 640×640×3s),POST /v1/library/video(text-to-video or one reference image), mp4/webm ingest +<video>poster cards, a motion quality gate (start/mid/end frames), and the Video template card. Built → adversarially verified (composer-lane-reset blocker fixed) → merged. No new infra. 1b. 3D — ✅ SHIPPED (2026-07-18). Hunyuan3D-2.1 image-to-mesh on the GB10:_model3d_graph(ImageOnlyCheckpointLoader bundled model+CLIP+VAE + ModelSamplingAuraFlow — both non-obvious and load-bearing) →.glb,POST /v1/library/model3d(one reference image, no text-to-3D), glTF ingest, and an interactive<model-viewer>zoom card. Verified: a clean 3D garment mesh from a ghost-mannequin shot. No new infra. - Music / audio — engine-served (ACE-class TTA model on the fleet); audio cards with waveform + play; gate = audio model scoring. Renders are cheap; the flywheel (notes like "too much reverb") ports verbatim.
- Decks — different shape: a deck is a COMPOSITION of library assets + text. Editor surface = HTML deck editor (one slide = one canvas, assets dragged from the library rail); generation = per-slide image lanes plus a copilot outline pass. Export: PDF/PNG via the render fleet. This is where "Canva breadth" lands — and every slide image is already flywheel-backed.
- Documents / brand kits — templates + design-system chips already in the composer become enforceable: a brand kit is a Stack of approved assets + palette + type rules the copilot injects into every lane's conditioning.
Editor architecture
One editor shell (the existing Studio⇄Editor toggle), N surfaces. A surface
registers {contentType, mount(el, asset), save() -> asset}. The node editor
is the image surface; video gets trim/poster controls on the same shell; decks
get the slide canvas. No per-medium apps — the shell, chrome, chat, wallet,
and library rail are one implementation.
What we explicitly do NOT build
- No separate video/music/deck libraries, queues, or billing.
- No third-party embed editors — surfaces are ours, small, and composable.
- No lane ships without its judge: ungated generation is how catalogs rot.
Rollout
Video lane first (it proves the lane contract with zero new infra), then audio, then the deck surface. Each lane lands behind its template card — no flags, no env gates; a lane is either shipped whole or absent.
Open questions
- Deck export fidelity target (PDF print vs web share) — decides the renderer.
- Music licensing posture for generated audio in client work.
- Whether the copilot outline pass for decks runs on the org's zen router or the default session model.