Each clip in Stage 4 hosts the shared media engine: a Higgsfield-only ProviderModelPicker, an editable composed video prompt (motion · camera · subject for a ~6s clip), a guide, and refine-with-history. The pink ClipCard is the studio chrome; the violet panel inside is the engine — a deliberate, contained seam. The composed / refined prompt pins to a per-clip snapshot so Generate clip below runs exactly it.
This is the per-clip GeneratePanel in detail. Below is Clip 2's engine section — model picker → editable composed VIDEO prompt + guide → refine row with revertable history. It composes on the free text path; the takes grid & Generate-clip row stay unchanged below (coexistence) and spend the media credit on the pinned snapshot.
2
Clip 2 · Walking hero ~6s
Mid-reel beat · subject in motion through the market · 9:16 · pins to this clip's prompt snapshot
Frames Ingredients
Shared media engine · Clip 2
Media provider
higgsfield
Per-clip composed generation is Higgsfield-only — google-flow stays in the manual / DIY path.
composed on platform key
Higgsfield video model
Seedance 2.0 — multi-shot
Model picked for this clip → the engine below is active (no model-less auto-preview).
Composed clip prompt — Clip 2 (video)
Prompt guide — video
1.Describe one continuous ~6s beat — subject action, then the camera move (dolly / push-in / pan) as if directing a single take.
2.Call the motion & light explicitly — stride, breeze, golden-hour rake, depth of field — so Seedance renders movement, not a still.
3.Name the on-brand assets (linen shirt, logo, Aanya) so the typed ingredients below are placed, not look-alikes.
Prompt composed by your AI provider; the clip is delivered by your Higgsfield connector. Composing & refining never spends a media credit.
Refine the prompt
Refining rewrites the text prompt on the free text path (basePrompt-chained) — it does not spend a media credit. Credits are only spent when you Generate clip below.
1more dynamic camera
2slower push-in on the final second
3keep the shirt logo legible mid-stride
Pinned to Clip 2's snapshot. The composed & refined prompt above is the exact text Generate clip runs — edits here re-pin the snapshot; the takes grid below renders from it.
Gated state — model-less clip — e.g. Clip 3, before a model is picked
3
Clip 3 · Packshot CTA
No video model picked yet → the engine is held disabled so the auto-preview can't fire model-less.
~5s
Pick a model for this clip first. The composed-prompt engine stays disabled — no model-less auto-preview, no wasted compose call — until you select a Higgsfield video model for Clip 3.
Shared media engine · disabled
Media provider
higgsfield
no model
Higgsfield video model
Select a model…
Composed clip prompt — Clip 3 (video)
Generate clip · takes — unchanged Studio chrome below, runs the pinned Clip 2 snapshot
Refine uses the free text path (basePrompt-chained) — not the heavy scene-rework job — so it never spends a media credit. Generate clip runs the pinned composed prompt on your Higgsfield connector and spends the media credit.