@koda-sl/baker-cli 0.305.0-dev.4e8ef1cd7 → 0.305.0-dev.a04fd407d

package/dist/cli.js

@@ -16,6 +16,7 @@ import {
   RunAbortedError,
   SEEDANCE_DURATIONS,
   SEEDANCE_PROFILE,
+  SPOKEN_CLIP_DURATION_S,
   SPOKEN_WORDS_PER_SECOND,
   VIDEO_GENERATE_MODELS,
   ValidationError,
@@ -65,7 +66,7 @@ import {
   ulid,
   validateCanvasDeep,
   ytDlpBlockSignal
-} from "./chunk-7TL4VPK5.js";
+} from "./chunk-ZNN6ZHDS.js";
 import {
   csvOrJson,
   daysAgoIso,
@@ -34098,6 +34099,27 @@ function joinDialogueTexts(texts) {
   }
   return out.join(" ");
 }
+function foldUnshownPhrases(phrases, keepRenderedVoice) {
+  if (!keepRenderedVoice) return [...phrases];
+  const claimed = /* @__PURE__ */ new Set();
+  const kept = [];
+  for (const phrase of phrases) {
+    const available = phrase.shownScenes.filter((scene) => !claimed.has(scene));
+    if (phrase.presenterShown && available.length > 0) {
+      for (const scene of available) claimed.add(scene);
+      kept.push(phrase);
+      continue;
+    }
+    const host = [...kept].reverse().find((candidate) => candidate.speaker === phrase.speaker);
+    if (!host) {
+      kept.push(phrase);
+      continue;
+    }
+    host.text = joinDialogueTexts([host.text, phrase.text]);
+    host.end_s = Math.max(host.end_s, phrase.end_s);
+  }
+  return kept;
+}
 function collapseVoiceover(blueprint) {
   const casts = castIdSet(blueprint);
   const cameraOn = onCameraDialogue(blueprint);
@@ -34660,7 +34682,7 @@ function buildTimeline(blueprint, slots, opts, nodes) {
   const phrases = buildPhrases(blueprint, canonical2, compositeScenes, presenterPresent, presentStrict);
   const usedVoIds = /* @__PURE__ */ new Set();
   const claimed = /* @__PURE__ */ new Set();
-  phrases.forEach((phrase, k) => {
+  foldUnshownPhrases(phrases, env.keepRenderedVoice === true).forEach((phrase, k) => {
     const voiceNode = ensureVoiceNode(phrase.speaker);
     const available = phrase.shownScenes.filter((s) => !claimed.has(s));
     if (phrase.presenterShown && available.length > 0) {
@@ -38557,6 +38579,7 @@ var CURRENCY_BY_MARKET = {
   "the united states": "$ (US dollars)"
 };
 var MARKERS_TO_NAME_A_PLACE = 3;
+var WORDS_PER_SECOND2 = 3.06;
 var MARKET_BY_LANGUAGE = {
   es: "Spain",
   "es-es": "Spain",
@@ -38629,9 +38652,10 @@ function marketFor(spec) {
 }
 function beatSeconds(line, videoModel) {
   const words2 = line.trim().split(/\s+/).length;
-  const spoken = words2 / 3.06;
+  const spoken = words2 / WORDS_PER_SECOND2;
   const allowed = (videoModel ? clipDurationsFor(videoModel) : void 0) ?? [...SEEDANCE_DURATIONS].sort((a, b) => a - b);
-  return allowed.find((d) => d >= spoken) ?? allowed[allowed.length - 1];
+  const wanted = Math.max(spoken, SPOKEN_CLIP_DURATION_S);
+  return allowed.find((d) => d >= wanted) ?? allowed[allowed.length - 1];
 }
 var BRAND_PLATE = "A solid flat brand colour plate \u2014 an end card. No people, no objects, no scenery, no text of any kind.";
 var ASKS_FOR_LEGIBLE = /\b(focus on the (number|amount|total|price|figure|text)|readable|legible|clearly (shows?|visible)|highlighted (number|amount|total)|showing the (number|amount|total|price))\b/i;
@@ -38861,6 +38885,22 @@ function adSpecToBlueprint(spec, opts = {}) {
     scenes
   };
 }
+function beatsTooShortToPlay(spec) {
+  return spec.beats.map((beat, i) => ({ i, spoken: beat.say.trim().split(/\s+/).filter(Boolean).length / WORDS_PER_SECOND2 })).filter(({ spoken }) => spoken < SPOKEN_CLIP_DURATION_S).map(({ i }) => i + 1);
+}
+function withoutCastName(text2, names) {
+  let out = text2;
+  for (const name of names) {
+    const token = name.trim();
+    if (token.length < 3) continue;
+    const pattern = new RegExp(`(?<![\\p{L}\\p{N}])${escapeForPattern(token)}(?![\\p{L}\\p{N}])`, "giu");
+    out = out.replace(pattern, "the person in the reference");
+  }
+  return out;
+}
+function escapeForPattern(value) {
+  return value.replace(/[.*+?^${}()|[\]\\]/g, "\\$&");
+}
 
 // src/engine/scaffold/ad-brand-overlay.ts
 var markFallback = (name) => name.trim() ? ` onerror="this.replaceWith(Object.assign(document.createElement('span'),{className:this.className+' brand-wordmark',textContent:${JSON.stringify(name.trim())}}))"` : "";
@@ -39362,6 +39402,13 @@ var scaffoldAdCommand = defineCommand111({
       process.exit(1);
       return;
     }
+    if (handle) {
+      const names = [handle, avatar?.name].filter((value) => Boolean(value));
+      spec.data.beats = spec.data.beats.map((beat) => ({ ...beat, show: withoutCastName(beat.show, names) }));
+      if (spec.data.cast?.description) {
+        spec.data.cast = { ...spec.data.cast, description: withoutCastName(spec.data.cast.description, names) };
+      }
+    }
     const blueprint = adSpecToBlueprint(spec.data, { avatarAccent: avatar?.profile?.accent, videoModel: chosenVideoModel });
     const slug = args.slug ?? path24.basename(specPath).replace(/\.[^.]+$/, "");
     const outPath = args.out ?? (args.slug ? path24.join("src/creatives", slug, `${slug}.canvas.json`) : path24.join(path24.dirname(specPath), `${slug}.canvas.json`));
@@ -39453,6 +39500,9 @@ var scaffoldAdCommand = defineCommand111({
       ok: true,
       data: { canvas: outPath, beats: spec.data.beats.length, slug },
       hints: [
+        // Before anything else: a beat too short to play is a beat to merge, and it is
+        // free to fix here and expensive to discover in a finished ad.
+        ...shortBeatHint(beatsTooShortToPlay(spec.data), spec.data.beats.length),
         ...noBrand ? [
           "Built with NO brand mark and no closing card, because you passed --no-brand. Say that in your reply: the user is getting an unbranded ad and should know it, not discover it. When they do have a logo, put it in `src/brand/logos/` and re-run without the flag."
         ] : [],
@@ -39511,6 +39561,12 @@ var scaffoldAdCommand = defineCommand111({
     });
   }
 });
+function shortBeatHint(short, total) {
+  if (short.length === 0) return [];
+  return [
+    `Beat${short.length === 1 ? "" : "s"} ${short.join(", ")} of ${total} ${short.length === 1 ? "is" : "are"} under ${SPOKEN_CLIP_DURATION_S}s \u2014 a beat lasts exactly as long as its line takes to read, so a two-word line is a three-second shot, and a run of those plays as a slideshow rather than an ad. Do NOT pad them: that buys seconds of somebody standing still. Merge them, so each beat carries a whole thought \u2014 an ad of this length usually wants two or three beats, not six. Edit the spec and re-run this command.`
+  ];
+}
 
 // src/commands/canvas/scaffold-video.ts
 import { access as access2, cp as cp2, mkdir as mkdir8, readFile as readFile19, rm as rm7, writeFile as writeFile10 } from "fs/promises";
@@ -57795,23 +57851,23 @@ function uploadFailure(path47) {
 }
 
 // src/commands/studio/spokenPace.ts
-var WORDS_PER_SECOND2 = SPOKEN_WORDS_PER_SECOND;
+var WORDS_PER_SECOND3 = SPOKEN_WORDS_PER_SECOND;
 var spokenWords = spokenWordsIn;
 function wordsThatFit(seconds) {
-  return Math.floor(seconds * WORDS_PER_SECOND2);
+  return Math.floor(seconds * WORDS_PER_SECOND3);
 }
 function spokenPaceHint(prompt, seconds, model) {
   const words2 = spokenWords(prompt);
   if (words2 === 0) return null;
   const split = model ? splitScriptFor(words2, model) : null;
   if (split) {
-    return `The line is ${words2} words \u2014 about ${Math.ceil(words2 / WORDS_PER_SECOND2)}s of speech \u2014 and ${model} renders at most ${split.maxSeconds}s (~${Math.floor(split.maxSeconds * WORDS_PER_SECOND2)} words). It will not refuse and it will not cut: it will RUSH, swallowing syllables. Split the script into ${split.clips} takes of roughly ${split.wordsPerClip} words each, rendered separately and cut together \u2014 break it where the thought breaks, not mid-sentence. Or move to a model that renders longer (bytedance/seedance-2.5 goes to 30s).`;
+    return `The line is ${words2} words \u2014 about ${Math.ceil(words2 / WORDS_PER_SECOND3)}s of speech \u2014 and ${model} renders at most ${split.maxSeconds}s (~${Math.floor(split.maxSeconds * WORDS_PER_SECOND3)} words). It will not refuse and it will not cut: it will RUSH, swallowing syllables. Split the script into ${split.clips} takes of roughly ${split.wordsPerClip} words each, rendered separately and cut together \u2014 break it where the thought breaks, not mid-sentence. Or move to a model that renders longer (bytedance/seedance-2.5 goes to 30s).`;
   }
   if (!seconds) return null;
   const fits = wordsThatFit(seconds);
   if (words2 <= fits) return null;
   const rate = (words2 / seconds).toFixed(1);
-  return `The line is ${words2} words and this clip is ${seconds}s \u2014 that is ${rate} words a second against a natural ${WORDS_PER_SECOND2}. The model will not run over, it will RUSH: syllables swallowed, endings clipped, and the take unusable however good the picture. Cut the line to about ${fits} words, or give it the length it needs (${Math.ceil(words2 / WORDS_PER_SECOND2)}s). Do not re-render the same line at the same length and hope.`;
+  return `The line is ${words2} words and this clip is ${seconds}s \u2014 that is ${rate} words a second against a natural ${WORDS_PER_SECOND3}. The model will not run over, it will RUSH: syllables swallowed, endings clipped, and the take unusable however good the picture. Cut the line to about ${fits} words, or give it the length it needs (${Math.ceil(words2 / WORDS_PER_SECOND3)}s). Do not re-render the same line at the same length and hope.`;
 }
 
 // src/commands/studio/animate.ts