| name | prompt-to-asset-routing |
| description | Route a creative brief to the right AI generator across images (photoreal, illustration, vector/SVG, text-in-image), video, 3D, and audio — plus the editing round-trips (inpaint, outpaint, background-removal, upscale) that beat a fresh generation. License-first, Grok-lean for images where competitive, draft-vs-final cost tiering. Provider/price specifics [verify-at-use]; every price [unverified]. |
Prompt-to-Asset Routing
The first decision: given a brief, which generator, which model, and text-to-image vs an editing round-trip. Images are the v1-deep lane; video/3D/audio are routed. License gate first, then aesthetics.
Generative-media judgment, landscape is volatile. Every provider/model/price is [verify-at-use]; every price is [unverified — confirm on provider pricing page]. Read the matrix, never hard-code a provider.
Workflow
- Traverse the modality-depth tree (
../../knowledge/generation-decision-trees.md). What is the asset — still image, hero/ambient video, 3D object, audio? Then pick depth.
- License-first routing (the license-first→Grok-lean→fallback tree). Is this a client site? Is
indemnity_required set? A FLUX-dev asset is worthless on a client site (non-commercial). Route to a Firefly-class indemnified provider when indemnity is required; Grok/xAI where competitive (flag: no IP indemnity); fallback chain otherwise.
- Draft-vs-final tiering. Iterate on a cheap draft model to lock composition + brand, then spend on one premium final render. Log spend with
gen-budget.py.
- Prefer an editing round-trip over a regeneration when a brand-locked asset exists.
- Hand off. Web-optimization →
web-optimization-pipeline; license/provenance → license-and-provenance-ledger; brand-style setup → brand-conditioned-generation; the ship gate → curation-and-accessibility-gate.
Images — the five capabilities (v1-deep)
| Capability | Route to (matrix) | Note |
|---|
| Photoreal | Grok (default where competitive) / Imagen 4 / FLUX.2 [pro] via BFL API | Never FLUX-dev on a client site |
| Illustration | Recraft / Firefly / Ideogram | Style-reference for brand consistency |
| Vector / SVG | Recraft V3/V4 (native SVG) | True vector for logo/icon — infinite scale, tiny, crisp |
| Text-in-image | Ideogram 3.0 (best legibility) — but still overlay real type | Models garble text; overlay HTML/SVG for the web |
| Brand-styled | Recraft brand-style upload (3–10 refs) + post-overlay exact hex | See brand-conditioned-generation |
Images — the four editing round-trips (first-class)
| Round-trip | When | Why it beats a fresh generation |
|---|
| Inpaint | Fix/replace a region (remove an artifact, swap an element) | Keeps the rest of the brand-locked composition intact |
| Outpaint | Extend the canvas (wider hero, new aspect ratio) | Reuse an approved image instead of re-rolling |
| Background-removal | Product cut-outs, transparent overlays | Deterministic, cheap, no re-generation |
| Upscale | Ship a small draft at hero resolution | Draft cheap, upscale the chosen one |
Rule: if a brand-locked asset already exists, reach for a round-trip before spending on a new generation.
Video / 3D / audio (routed)
- Video → Veo/Kling/Runway/Sora class (commercial use plan-dependent
[verify-at-use]); the real embed pipeline is web-optimization-pipeline.
- 3D → Meshy/Rodin/TRELLIS → glTF/GLB;
<model-viewer> embed.
- Audio → capable API via fal; no autoplay audio on the web.
Anti-patterns
- Routing on aesthetics before the license (the FLUX-dev trap).
- A fresh generation when an inpaint/outpaint/upscale round-trip would do.
- Baking text into an image instead of overlaying real type.
- Hard-coding a provider instead of reading the matrix.
See also
