@koda-sl/baker-cli 0.133.1 → 0.133.3

package/dist/cli.js

@@ -34,7 +34,7 @@ import {
   toModelSafeImage,
   ulid,
   validateCanvasDeep
-} from "./chunk-Q4UI5JZN.js";
+} from "./chunk-CIF62V7L.js";
 import {
   csvOrJson,
   daysAgoIso,
@@ -20631,20 +20631,6 @@ function parseBudget(raw) {
 }
 function resolveModels2(args) {
   const pick = (flag, kind, fallback) => args[flag] ? String(args[flag]) : resolveModel2(kind, fallback);
-  let videoModel;
-  let videoRouteReason;
-  if (args["video-model"]) {
-    videoModel = resolveModel2("video_generate", String(args["video-model"]));
-  } else {
-    const route = routeVideoModel({
-      hasRealFace: Boolean(args["real-face"]),
-      needsIdentity: Boolean(args.identity),
-      needsMotionTransfer: Boolean(args["motion-transfer"]),
-      budget: parseBudget(args.budget)
-    });
-    videoModel = resolveModel2("video_generate", route.model);
-    videoRouteReason = route.reason;
-  }
   return {
     // Flash by default: ~2-3× faster per call than Pro, which keeps each
     // per-scene deconstruct step well inside the action time budget and turns a
@@ -20655,11 +20641,29 @@ function resolveModels2(args) {
     // Default to the strongest image model (matches the static-ad scaffold); the
     // frame generators need the most faithful text/identity reproduction. Override
     // with --image-model for a cheaper/faster pass.
-    imageModel: pick("image-model", "image_generate", "openai/gpt-5.4-image-2"),
-    videoModel: videoModel || DEFAULT_VIDEO_GENERATE_MODEL,
-    videoRouteReason
+    imageModel: pick("image-model", "image_generate", "openai/gpt-5.4-image-2")
   };
 }
+function hasPhotorealCast(elements) {
+  return Array.isArray(elements) && elements.some((e) => {
+    const type = typeof e === "object" && e !== null ? e.type : void 0;
+    return type === "person" || type === "animal";
+  });
+}
+function resolveVideoModel(args, elements) {
+  if (args["video-model"]) {
+    return { videoModel: resolveModel2("video_generate", String(args["video-model"])) };
+  }
+  const photorealCast = hasPhotorealCast(elements);
+  const route = routeVideoModel({
+    hasRealFace: Boolean(args["real-face"]) || photorealCast,
+    needsIdentity: Boolean(args.identity),
+    needsMotionTransfer: Boolean(args["motion-transfer"]),
+    budget: parseBudget(args.budget)
+  });
+  const reason = photorealCast && !args["real-face"] ? `${route.reason} \u2014 detected a photoreal on-camera cast, which Seedance's real-person filter would block` : route.reason;
+  return { videoModel: resolveModel2("video_generate", route.model) || DEFAULT_VIDEO_GENERATE_MODEL, videoRouteReason: reason };
+}
 function buildDeconstructCanvas(videoPath, deconstructModel, opts) {
   const deconstructParams = { model: deconstructModel, mode: "full" };
   if (typeof opts.maxScenes === "number") deconstructParams.max_scenes = opts.maxScenes;
@@ -20780,7 +20784,7 @@ var scaffoldVideoCommand = defineCommand92({
     "video-model": { type: "string", description: "Override the video_generate model id for clips (skips the scored router)" },
     "real-face": {
       type: "boolean",
-      description: "Brief needs a real human likeness preserved \u2192 the router picks Veo (dodges the ByteDance real-person filter)."
+      description: "The ad has a photoreal human face on camera \u2192 the router picks Veo (dodges the ByteDance real-person filter). Set this for ANY photoreal presenter/creator/testimonial cast \u2014 the filter blocks AI-GENERATED photoreal faces too, not just real ones."
     },
     "motion-transfer": {
       type: "boolean",
@@ -20839,7 +20843,7 @@ var scaffoldVideoCommand = defineCommand92({
 `
       );
     }
-    const { deconstructModel, selectModel, imageModel, videoModel, videoRouteReason } = resolveModels2(args);
+    const { deconstructModel, selectModel, imageModel } = resolveModels2(args);
     const shotThreshold = args["shot-threshold"] ? Number(args["shot-threshold"]) : void 0;
     const shotCuts = await detectShotCutsBestEffort(videoPath, shotThreshold);
     const deconstructCanvas = buildDeconstructCanvas(videoPath, deconstructModel, {
@@ -20849,6 +20853,11 @@ var scaffoldVideoCommand = defineCommand92({
       shotCuts
     });
     const { blueprint, elements, transcript, creditsSpent } = await runAnalysisPasses(deconstructCanvas, selectModel);
+    const { videoModel, videoRouteReason } = resolveVideoModel(args, elements);
+    if (videoRouteReason && !args["video-model"] && !args["real-face"] && hasPhotorealCast(elements)) {
+      process.stderr.write(`\u{1F3AC} Photoreal on-camera cast detected \u2192 generating clips on ${videoModel} (dodges Seedance's real-person filter).
+`);
+    }
     await mkdir6(outDir, { recursive: true });
     const annotated = annotateBlueprintWithElements(blueprint, elements);
     await writeSceneFiles(outDir, annotated);
@@ -21018,12 +21027,12 @@ var scaffoldVideoCommand = defineCommand92({
             // on-brand color. The overlays inherit the reference's motion unless the
             // operator restyles easing/transition/accent from the brand's motion language.
             prompt_discipline: "Rewrite each scene with the six-decision structure (Route / Spec / Beats / Copy / Technique / Negatives): quote exact copy, give every beat a timestamp, and name the moves. Pull easing, transition (cut vs fade), pacing, and the accent moment from BRAND.md \xA7 Brand in Motion so the overlays move like THIS brand \u2014 not the reference's. If BRAND.md has no Brand-in-Motion section, seed it with brand-build first. See the creative-canvas-ads skill: references/prompt-anatomy.md + references/hyperframes/.",
-            // A photoreal on-camera person/animal on Seedance can trip ByteDance's
-            // real-person-likeness filter (422 content_policy_blocked, NON-retryable — no
-            // prompt reframe clears it). Surface the escape BEFORE the billed run so a
-            // face-heavy ad isn't discovered broken mid-render.
+            // Photoreal on-camera casts auto-route to Veo above, so this only fires
+            // when `--video-model bytedance/seedance-2.0` was EXPLICITLY forced despite
+            // a person/animal cast — overriding that safety route. Warn that every such
+            // clip will hit ByteDance's real-person filter and burn a wasted attempt.
             ...report.elements.some((e) => e.type === "person" || e.type === "animal") && /seedance/i.test(videoModel) ? {
-              content_policy_risk: "This ad has a photoreal on-camera cast generating on Seedance. ByteDance's real-person-likeness filter can reject a photoreal AI face with a NON-retryable 422 (content_policy_blocked) \u2014 no prompt change clears it. If clips fail that way, regenerate on Veo (re-run with `--video-model google/veo-3.1-fast`) or make the frame less photoreal."
+              content_policy_risk: "You FORCED Seedance (`--video-model bytedance/seedance-2.0`) on an ad with a photoreal on-camera cast. ByteDance's real-person filter will reject those clips (content_policy_blocked, E005) \u2014 no prompt change clears it. Each clip auto-recovers on Veo at render time so the run still completes, but you burn a wasted Seedance attempt per clip. Drop `--video-model` to let the scaffold route the cast to Veo up front, or make the frames less photorealistic."
             } : {},
             note: "Drop ONE real source image at each el_* [TODO] (reused across every frame that element appears in), confirm each voice_select casting, then `baker canvas validate` and `baker canvas run`. Running generates many billed image/video/audio assets \u2014 it is not free."
           }