@slatesvideo/shared 0.4.8 → 0.5.1

package/dist/prompts/model-facts.js

@@ -12,6 +12,30 @@ export const MODEL_FACTS = [
         maxIngredients: null,
         notes: 'Default image model. 14 refs hard cap (10 object + 4 character). Brief it like a creative director, not tag soup. No negativePrompt field — use positive reframing. Best image start-frame for legible text. Knowledge cutoff Jan 2025.',
     },
+    {
+        id: 'nano-banana-2-lite',
+        label: 'Nano Banana 2 Lite',
+        kind: 'image',
+        maxRefImages: 4,
+        maxIngredients: null,
+        notes: 'FAST/DRAFT image tier — ~half the price of NB2 full, ~2.7× faster, 1K output ONLY. Same Gemini content filter as NB2. Route here for iteration volume and drafts where 1K is fine; keep NB2 full for final 2K/4K. Character consistency + legible text hold up.',
+    },
+    {
+        id: 'nano-banana-pro',
+        label: 'Nano Banana Pro',
+        kind: 'image',
+        maxRefImages: 14,
+        maxIngredients: null,
+        notes: 'HERO-FRAME / typography PREMIUM image tier (Gemini 3 Pro backbone; ~2× NB2 price). NB2 ≈ 95% of Pro — route here only when spatial composition, cinematic lighting/skin, fine typography-in-scene, or deep multi-element reasoning must be perfect. Up to 14 reference images (character locking, multi-subject fusion). Native 16:9 + 4K.',
+    },
+    {
+        id: 'gpt-image-2',
+        label: 'GPT Image 2',
+        kind: 'image',
+        maxRefImages: 10,
+        maxIngredients: null,
+        notes: 'TEXT/DIAGRAM/PANEL king — near-perfect character-level text, ordered panels, exact placement (~3s gens). Route here for character sheets, shot grids, and text-bearing panels. Quality tiers: medium (default, the value seat — half NB2 price at 1080p) / high (~4×, max text precision). Third filter regime (OpenAI moderate). 4K is API-only — even paid ChatGPT can\'t render it. Photoreal/character-locked/edit-heavy → Banana line instead.',
+    },
     {
         id: 'flux-2-max',
         label: 'FLUX.2 Max',
@@ -34,7 +58,7 @@ export const MODEL_FACTS = [
         kind: 'video',
         maxRefImages: null,
         maxIngredients: 9, // ingredient images per video gen
-        notes: 'PREMIUM video tier — route here the moment physics, effects, destruction, or scale matter, and for hero shots. VIDEO-ONLY: cannot generate standalone images (use NB2/FLUX.2/Seedream for those). Up to 9 ingredient images. Strong I2V / own-footage restyle. Native 4K.',
+        notes: 'PREMIUM video tier — route here the moment physics, effects, destruction, or scale matter, and for hero shots. VIDEO-ONLY: cannot generate standalone images (use NB2/FLUX.2/Seedream for those). Up to 9 ingredient images. Strong I2V / own-footage restyle. Native 4K. Also the PREMIUM engine inside the Motion Transfer and Lip Sync tools (single-pass: driving video / dialogue are native conditioning signals — better motion fidelity, natural speech, voice cloned from a video source; video references bill input+output seconds).',
     },
     {
         id: 'kling-v3',
@@ -42,7 +66,15 @@ export const MODEL_FACTS = [
         kind: 'video',
         maxRefImages: null,
         maxIngredients: 4,
-        notes: 'DEFAULT general-purpose video model — cost-effective, strong start-frame adherence (identity/layout/text), acting, dialogue, lip-sync, any aspect ratio. Escalate to Seedance for physics.',
+        notes: 'DEFAULT general-purpose video model — cost-effective, strong start-frame adherence (identity/layout/text), acting, dialogue, lip-sync, any aspect ratio. Escalate to Seedance for physics. In the Motion Transfer / Lip Sync tools, Kling (MC / lip-sync / avatar) is the cheap utility lane; Seedance is the premium single-pass lane.',
+    },
+    {
+        id: 'kling-v3-edit',
+        label: 'Kling O3 Video Edit',
+        kind: 'video',
+        maxRefImages: null,
+        maxIngredients: 4, // combined subject elements + style refs per edit
+        notes: 'VIDEO-TO-VIDEO EDIT (default edit tool) — takes an EXISTING 3–15s clip and changes only what the prompt names: character swap, environment change, style transfer. Original motion/camera/audio preserved. One pass, no masking. Use to FIX a 90%-right clip instead of re-rolling it, or to AI-edit the user\'s own footage. Elements (@ElementN = frontal + angles) lock subject identity; max 4 combined refs. Billed per second of output (≈ clip length, rounded up). Seedance edit/relocate is the alternative for style-transfer-heavy jobs.',
     },
     {
         id: 'veo-3.1',

package/dist/prompts/reference-composer.d.ts

@@ -43,4 +43,39 @@ export interface ComposeOptions {
     startVideoNumber?: number;
 }
 export declare function composeReferences(rawPrompt: string, groups: ReferenceGroup[], opts?: ComposeOptions): ComposedReferences;
+export interface KlingEditElement {
+    /** Frontal image path/URL (element primary view) */
+    frontal: string;
+    /** Up to 3 additional angle images */
+    angles: string[];
+    /** Display name (for logging/echo) */
+    name: string;
+}
+export interface KlingEditComposition {
+    /** Prompt in Kling's edit notation: @Video1 (source), @ElementN, @ImageN */
+    prompt: string;
+    /** Subject elements in cited order (@Element1..) */
+    elements: KlingEditElement[];
+    /** Style/appearance reference image paths in cited order (@Image1..) */
+    styleImages: string[];
+}
+/** fal cap: max 4 combined element + style-image references per edit request. */
+export declare const KLING_EDIT_MAX_REFS = 4;
+/**
+ * Compose references for the Kling O3 edit endpoint. Unlike the "image N"
+ * naming every other model parses, Kling's edit endpoint has its OWN official
+ * notation: `@Video1` (the source clip), `@Element1..` (subjects, frontal +
+ * angle images), `@Image1..` (style/appearance refs). This adapter translates
+ * the one-ordered-list groups into that notation — the same prompt-is-law
+ * principle, different citation tokens.
+ *
+ * - character/environment groups → elements (media[0] = frontal, rest = angles,
+ *   max 3 angles per fal's schema)
+ * - style/pinned image groups → @ImageN style refs
+ * - video groups are ignored here — the source clip is transported separately
+ *   and is always @Video1
+ * - combined element+image count is capped at KLING_EDIT_MAX_REFS (style refs
+ *   trimmed first — subjects are the feature)
+ */
+export declare function composeKlingEdit(rawPrompt: string, groups: ReferenceGroup[]): KlingEditComposition;
 //# sourceMappingURL=reference-composer.d.ts.map

package/dist/prompts/reference-composer.js

@@ -172,4 +172,85 @@ export function composeReferences(rawPrompt, groups, opts = {}) {
         orderedVideoPaths,
     };
 }
+/** fal cap: max 4 combined element + style-image references per edit request. */
+export const KLING_EDIT_MAX_REFS = 4;
+/**
+ * Compose references for the Kling O3 edit endpoint. Unlike the "image N"
+ * naming every other model parses, Kling's edit endpoint has its OWN official
+ * notation: `@Video1` (the source clip), `@Element1..` (subjects, frontal +
+ * angle images), `@Image1..` (style/appearance refs). This adapter translates
+ * the one-ordered-list groups into that notation — the same prompt-is-law
+ * principle, different citation tokens.
+ *
+ * - character/environment groups → elements (media[0] = frontal, rest = angles,
+ *   max 3 angles per fal's schema)
+ * - style/pinned image groups → @ImageN style refs
+ * - video groups are ignored here — the source clip is transported separately
+ *   and is always @Video1
+ * - combined element+image count is capped at KLING_EDIT_MAX_REFS (style refs
+ *   trimmed first — subjects are the feature)
+ */
+export function composeKlingEdit(rawPrompt, groups) {
+    const elements = [];
+    const styleImages = [];
+    const styleNums = [];
+    let body = rawPrompt;
+    // Subjects first — they own the @ElementN numbering.
+    const subjectGroups = groups.filter((g) => (g.kind === 'character' || g.kind === 'environment') && g.media.some((m) => m.mediaKind === 'image'));
+    for (const g of subjectGroups) {
+        if (elements.length >= KLING_EDIT_MAX_REFS)
+            break;
+        const imgs = g.media.filter((m) => m.mediaKind === 'image').map((m) => m.path);
+        if (imgs.length === 0)
+            continue;
+        const n = elements.length + 1;
+        elements.push({ frontal: imgs[0], angles: imgs.slice(1, 4), name: g.name });
+        if (g.token) {
+            // Replace every @token occurrence with the element citation.
+            const escaped = g.token.replace(/[.*+?^${}()|[\]\\]/g, '\\$&');
+            const re = new RegExp(`${escaped}\\b`, 'gi');
+            if (re.test(body)) {
+                body = body.replace(re, `@Element${n}`);
+            }
+            else {
+                body = `${body}\n@Element${n} is ${g.name}.`;
+            }
+        }
+        else {
+            body = `${body}\n@Element${n} is ${g.name}.`;
+        }
+    }
+    // Style / pinned refs take the remaining slots as @ImageN.
+    for (const g of groups) {
+        if (g.kind !== 'style' && g.kind !== 'pinned')
+            continue;
+        for (const m of g.media) {
+            if (m.mediaKind !== 'image')
+                continue;
+            if (elements.length + styleImages.length >= KLING_EDIT_MAX_REFS)
+                break;
+            styleImages.push(m.path);
+            const n = styleImages.length;
+            if (g.kind === 'style')
+                styleNums.push(n);
+            if (g.token) {
+                const escaped = g.token.replace(/[.*+?^${}()|[\]\\]/g, '\\$&');
+                body = body.replace(new RegExp(`${escaped}\\b`, 'gi'), `@Image${n}`);
+            }
+        }
+    }
+    if (styleNums.length > 0) {
+        const cites = styleNums.map((n) => `@Image${n}`).join(' and ');
+        body = `${body}\nApply the visual style of ${cites}.`;
+    }
+    // The source clip is always @Video1 — anchor the instruction to it.
+    if (!/@Video1\b/i.test(body)) {
+        body = `Edit @Video1: ${body}`;
+    }
+    return {
+        prompt: body.replace(/[ \t]{2,}/g, ' ').trim(),
+        elements,
+        styleImages,
+    };
+}
 //# sourceMappingURL=reference-composer.js.map