@gentbajko/slopify 3.0.11 → 3.0.12

package/dist/slices/rebuild/recipe-visual.js

@@ -3,6 +3,7 @@ import { usesShortMode } from "../admission/short-mode.js";
 import { render } from "../admission/substitute.js";
 import { withoutScene, withScene } from "../images/scenes.js";
 import { usesAmbientBed } from "../video/ambient-bed.js";
+import { lookWait, withLooks } from "./recipe-appearance.js";
 import { castFor } from "./recipe-cast.js";
 import { noMaster } from "./recipe-loudness.js";
 import { recipe, resourceIdentity, } from "./recipe-model.js";
@@ -18,7 +19,9 @@ timing = null,
 // Level the volume: the levelled narration and the master (`recipe-loudness.ts`).
 master = noMaster, 
 // Each image's own scene, for the images whose prompt asks for one (`recipe-scenes.ts`).
-scenes) {
+scenes, 
+// How the subject and characters look, for prompts with `{{Appearance}}`.
+appearance) {
     const recipes = [];
     const imageKeys = config.sources.images === "off" ? [] : content.imageOrder;
     for (const key of imageKeys) {
@@ -30,15 +33,37 @@ scenes) {
             : content.promptTemplates[image.templateKey];
         const at = scenes?.keys.indexOf(key) ?? -1;
         if (scenes !== undefined && at !== -1 && image.source === "generate") {
-            recipes.push(sceneImage(config, content, key, raw ?? image.prompt ?? "", scenes, at, reference));
+            recipes.push(sceneImage(config, content, key, raw ?? image.prompt ?? "", scenes, at, reference, appearance));
             continue;
         }
         // With Scenes from the article off, a prompt's `{{Scene}}` line is left out.
-        const prompt = raw === undefined || raw === null
+        const written = raw === undefined || raw === null
             ? image.prompt === null || image.prompt === undefined
                 ? image.prompt
                 : withoutScene(image.prompt)
             : render(withoutScene(raw), config.values);
+        const prompt = written === null || written === undefined || image.source !== "generate"
+            ? written
+            : withLooks(written, appearance);
+        // Waiting for the looks: the same wait an image with a scene has, on the lookup instead.
+        if (prompt === null && written !== null && written !== undefined) {
+            const wait = lookWait(written, appearance);
+            recipes.push(recipe({ content }, `image:${key}`, "images", {
+                kind: "deferred",
+                version: 1,
+                operation: "image-scene",
+                template: [
+                    ...wait.template,
+                    written,
+                    config.images?.provider ?? null,
+                    config.images?.model ?? null,
+                    config.format,
+                    ...(config.images?.thinking === undefined ? [] : [config.images.thinking]),
+                    ...(reference === undefined ? [] : [reference.input.fingerprint]),
+                ],
+            }, [...wait.dependsOn, ...(reference === undefined ? [] : [reference.key])]));
+            continue;
+        }
         recipes.push(recipe({ content }, `image:${key}`, "images", image.source === "provide"
             ? { kind: "provided", version: 1, assetId: image.assetId, semantic: null }
             : {
@@ -125,9 +150,12 @@ scenes) {
 // An image drawn from its own scene. Until the scenes are written it waits as a deferred
 // request that names the scenes step and the image's place among them, so writing them turns
 // it into its image request (materialization), and writing them again draws it again.
-function sceneImage(config, content, key, body, scenes, at, reference) {
+function sceneImage(config, content, key, body, scenes, at, reference, appearance) {
     const scene = scenes.scenes?.[at];
-    const prompt = scene === undefined ? undefined : withScene(render(body, config.values), scene);
+    const prompt = scene === undefined
+        ? undefined
+        : (withLooks(withScene(render(body, config.values), scene), appearance, scene) ?? undefined);
+    const wait = lookWait(body, appearance);
     return recipe({ content }, `image:${key}`, "images", prompt === undefined
         ? {
             kind: "deferred",
@@ -141,6 +169,7 @@ function sceneImage(config, content, key, body, scenes, at, reference) {
                 config.format,
                 ...(config.images?.thinking === undefined ? [] : [config.images.thinking]),
                 ...(reference === undefined ? [] : [reference.input.fingerprint]),
+                ...wait.template,
             ],
         }
         : {
@@ -151,7 +180,7 @@ function sceneImage(config, content, key, body, scenes, at, reference) {
             prompt,
             ...(reference === undefined ? {} : { reference: reference.input }),
             ...castField(castFor(config, prompt)),
-        }, [scenes.recipe.key, ...(reference === undefined ? [] : [reference.key])]);
+        }, [scenes.recipe.key, ...wait.dependsOn, ...(reference === undefined ? [] : [reference.key])]);
 }
 // What the ambient bed adds to the render's fingerprint: its settings and, for the user's own
 // file, the project asset it plays. Undefined without a bed.
@@ -170,7 +199,11 @@ export function ambientBedValues(config, content) {
 }
 export function thumbnailRecipes(context, textRecipes, 
 // The establishing image, when it is on and the thumbnail is drawn from it too.
-reference) {
+reference, 
+// Scenes from the article: a thumbnail prompt with `{{Scene}}` takes one per thumbnail.
+scenes, 
+// How the subject and characters look, for a prompt with `{{Appearance}}`.
+appearance) {
     const { config, content } = context;
     if (config.sources.thumbnail === "off")
         return [];
@@ -184,40 +217,65 @@ reference) {
             }, [], { unresolved: content.provided.thumbnail === undefined }),
         ];
     const promptRecipe = textRecipes.find((value) => value.key === "thumbnail:prompt");
-    const prompt = config.sources.thumbnail === "prompt_by_llm" && promptRecipe !== undefined
+    const written = config.sources.thumbnail === "prompt_by_llm" && promptRecipe !== undefined
         ? matchingText(context, promptRecipe, "prompt")
         : renderedPrompt(context, "thumbnailPrompt");
+    const sceneFor = (variant) => scenes === undefined || scenes.thumbnailCount === 0
+        ? undefined
+        : scenes.thumbnails?.[variant - 1];
+    const withScenes = scenes !== undefined && scenes.thumbnailCount > 0;
     const variants = Array.from({ length: thumbnailCountOf(config) }, (_, index) => index + 1);
-    return variants.map((variant) => recipe(context, thumbnailKey(variant), "thumbnail", prompt === null
-        ? {
-            kind: "deferred",
-            version: 1,
-            operation: "thumbnail-image",
-            template: [
-                promptRecipe?.fingerprint ?? null,
-                config.images?.provider ?? null,
-                config.images?.model ?? null,
-                config.format,
-                // Only what is in use, so a thumbnail made before these existed keeps its
-                // fingerprint.
-                ...(config.images?.thinking === undefined ? [] : [config.images.thinking]),
-                ...(reference === undefined ? [] : [reference.input.fingerprint]),
-                ...(variant === 1 ? [] : [["thumbnail-variant", variant]]),
-            ],
-        }
-        : {
-            kind: "image",
-            version: 1,
-            ...imageChoice(config),
-            aspect: config.format,
-            prompt: thumbnailVariantPrompt(prompt, variant),
-            ...(reference === undefined ? {} : { reference: reference.input }),
-            // The thumbnail stands for the whole video, like the establishing image.
-            ...castField(castFor(config, config.title, prompt)),
-        }, [
-        ...(promptRecipe === undefined ? [] : [promptRecipe.key]),
-        ...(reference === undefined ? [] : [reference.key]),
-    ]));
+    return variants.map((variant) => {
+        const scene = sceneFor(variant);
+        // With a scene the thumbnail waits for it; with the switch off a `{{Scene}}` line is left
+        // out, as the images do.
+        const sceneFilled = written === null
+            ? null
+            : withScenes
+                ? scene === undefined
+                    ? null
+                    : withScene(written, scene)
+                : withoutScene(written);
+        // A thumbnail stands for the subject, so it always has the subject's look.
+        const prompt = sceneFilled === null
+            ? null
+            : withLooks(sceneFilled, appearance, scene, { subjectAlways: true });
+        const wait = lookWait(written, appearance);
+        return recipe(context, thumbnailKey(variant), "thumbnail", prompt === null
+            ? {
+                kind: "deferred",
+                version: 1,
+                operation: "thumbnail-image",
+                template: [
+                    promptRecipe?.fingerprint ?? null,
+                    config.images?.provider ?? null,
+                    config.images?.model ?? null,
+                    config.format,
+                    // Only what is in use, so a thumbnail made before these existed keeps its
+                    // fingerprint.
+                    ...(config.images?.thinking === undefined ? [] : [config.images.thinking]),
+                    ...(reference === undefined ? [] : [reference.input.fingerprint]),
+                    ...(variant === 1 ? [] : [["thumbnail-variant", variant]]),
+                    ...(withScenes ? [["thumbnail-scene", scenes.recipe.fingerprint]] : []),
+                    ...wait.template,
+                ],
+            }
+            : {
+                kind: "image",
+                version: 1,
+                ...imageChoice(config),
+                aspect: config.format,
+                prompt: thumbnailVariantPrompt(prompt, variant),
+                ...(reference === undefined ? {} : { reference: reference.input }),
+                // The thumbnail stands for the whole video, like the establishing image.
+                ...castField(castFor(config, config.title, prompt)),
+            }, [
+            ...(promptRecipe === undefined ? [] : [promptRecipe.key]),
+            ...(reference === undefined ? [] : [reference.key]),
+            ...(withScenes ? [scenes.recipe.key] : []),
+            ...wait.dependsOn,
+        ]);
+    });
 }
 // The second and third thumbnails are the same prompt asked for another composition, so the
 // three test one idea against itself rather than three ideas. The first is the prompt as it

package/dist/slices/rebuild/recipe-work.js

@@ -281,7 +281,9 @@ export function priceRecipe(work, value, config) {
     // Scenes from the article: an image waiting for its scene is one image, and the scenes are
     // one call that reads the article (its length is known once the article is written).
     if (input.kind === "deferred" &&
-        (input.operation === "image-scene" || input.operation === "image-scenes")) {
+        (input.operation === "image-scene" ||
+            input.operation === "image-scenes" ||
+            input.operation === "image-appearance")) {
         const choice = config === undefined ? undefined : recipeProviderChoice(value, config);
         if (choice !== undefined && input.operation === "image-scene")
             return { kind: "image", stage: work.key, provider: choice.provider, model: choice.model };
@@ -292,7 +294,7 @@ export function priceRecipe(work, value, config) {
                 provider: choice.provider,
                 model: choice.model,
                 inputCharacters: 90000,
-                outputCharacters: 4000,
+                outputCharacters: input.operation === "image-appearance" ? 6000 : 4000,
                 detail: "The article is read when the step runs; its length is estimated. Retries are excluded.",
             };
     }

package/dist/slices/rebuild/runtime-provider.js

@@ -2,6 +2,7 @@ import { basename } from "node:path";
 import { referenceKey, thumbnailVariant } from "../admission/model.js";
 import { plainText } from "../article/plain.js";
 import { splitEndMatter } from "../article/split.js";
+import { checkAppearance } from "../images/appearance.js";
 import { checkScenes, sceneCountOf } from "../images/scenes.js";
 import { spokenPassage } from "../narration/describe.js";
 import { observeNarration } from "../narration/live.js";
@@ -15,6 +16,7 @@ import { outputPath } from "../storage/layout.js";
 import { probeDurationMs } from "../video/ffmpeg.js";
 import { parseAttribution } from "../voices/attribution.js";
 import { parseScript } from "../voices/script.js";
+import { imageAppearanceKey } from "./recipe-appearance.js";
 import { imageScenesKey } from "./recipe-scenes.js";
 import { executeArticleRequests } from "./runtime-article.js";
 import { imageCall } from "./runtime-image.js";
@@ -200,6 +202,11 @@ function checkAnswer(piece, answer) {
         const checked = checkScenes(answer.text, sceneCountOf(piece.input.messages) ?? 0);
         return checked.ok ? undefined : checked.reason;
     }
+    // The looks it looked up: at least the video's subject.
+    if (piece.key === imageAppearanceKey && piece.input.kind === "llm") {
+        const checked = checkAppearance(answer.text);
+        return checked.ok ? undefined : checked.reason;
+    }
     if (piece.key === "research:planner")
         return chaptersFrom(answer.text).length === 0
             ? "The AI model's research plan listed no chapters, so research could not go on. Use Try again; if it keeps happening, choose another model in the Providers section of Edit project."
@@ -226,6 +233,15 @@ async function publishText(deps, context, piece, answer) {
         ]), { scenes: checked.scenes, text: answer.text });
         return;
     }
+    if (piece.key === imageAppearanceKey && piece.input.kind === "llm") {
+        const checked = checkAppearance(answer.text);
+        if (!checked.ok)
+            throw new Error(checked.reason);
+        await publishResult(deps, context, piece, preparedTexts(deps, context, piece, [
+            ["instructions", "instructions.md", frozenInstructions(deps, context, piece)],
+        ]), { appearance: checked.appearance, text: answer.text });
+        return;
+    }
     if (piece.input.kind === "llm" && piece.input.preparation !== undefined) {
         const checked = validatePreparation(answer.text, piece.input.preparation.source);
         if (!checked.ok)

package/dist/slices/revisions/publication-rules.js

@@ -3,7 +3,7 @@ import { z } from "zod";
 import { workExists } from "../../kernel/runner/work-authority.js";
 import { pieceFile } from "../storage/reconcile.js";
 import { outputSchema } from "../storage/schema.js";
-import { outputsForRevision, piecesForRevision, revisionById } from "./repo.js";
+import { currentRevisionId, outputsForRevision, piecesForRevision, revisionById } from "./repo.js";
 import { stagePieceSchema } from "./schema.js";
 export function publicationAuthority(db, publication) {
     const { work, pieceId, publicationId } = publication;
@@ -42,6 +42,14 @@ export function validatePublication(db, publication, outputs, pieces) {
             .prepare("SELECT logical_fingerprint FROM revision_work_pieces WHERE id=? AND work_id=?")
             .get(publication.pieceId, publication.work.workId);
     const revision = revisionById(db, publication.work.projectId, publication.work.revisionId);
+    // The captions an export carries are the ones selected when it started. An edit while it ran
+    // (a thumbnail redo, say) makes a newer revision: captions made after that are selected there
+    // only, so the current revision counts as well as the work's own.
+    const head = currentRevisionId(db, publication.work.projectId);
+    const captionRevisions = [
+        publication.work.revisionId,
+        ...(head === undefined || head === publication.work.revisionId ? [] : [head]),
+    ];
     for (const row of outputs) {
         const retainedMember = ((authority?.key === "subtitles:files" &&
             ((row.workKey === "export:wav" && row.output.role === "audio_export") ||
@@ -51,12 +59,12 @@ export function validatePublication(db, publication, outputs, pieces) {
             ((authority?.key === "export:wav" || authority?.key === "export:video") &&
                 row.workKey === "subtitles:files" &&
                 ["subtitles_srt", "subtitles_vtt", "subtitle_ass", "subtitle_font"].includes(row.output.role))) &&
-            outputsForRevision(db, publication.work.projectId, publication.work.revisionId).some((old) => old.selected &&
+            captionRevisions.some((revisionId) => outputsForRevision(db, publication.work.projectId, revisionId).some((old) => old.selected &&
                 old.state === "ready" &&
                 old.workKey === row.workKey &&
                 old.output.role === row.output.role &&
                 old.assetId === row.asset.id &&
-                old.fingerprint === row.fingerprint);
+                old.fingerprint === row.fingerprint));
         if (!retainedMember &&
             row.fingerprint !== revision?.fingerprints[row.workKey] &&
             !(row.workKey === authority?.key &&

package/dist/tutorials/Play-Images.md

@@ -118,9 +118,19 @@ Where the scene goes:
 - A prompt with `{{Scene}}` gets the scene there, for example a line `Scene: {{Scene}}` under the prompt's first sentence. `{{Scene}}` is never a field to fill in: Slopify writes it.
 - A prompt without it gets the scene as a line of its own after its first paragraph, so any prompt works unchanged.
 - With the switch off, a line holding `{{Scene}}` is left out and the prompt is drawn as it always was.
+- A **thumbnail** prompt with `{{Scene}}` gets a scene from the same call: the article's most striking moment, the one that would make someone click, taken from anywhere in the article. With **Thumbnails** set to 3, each thumbnail gets a different moment, so the A/B test compares ideas as well as compositions. A thumbnail prompt without `{{Scene}}` is drawn as it always was.
 
 The style, composition and everything else in the prompt stay as written; only the scene changes from image to image. The scenes are written once the article exists, and each image waits for them. Switching it on or off in **Edit project** remakes the images.
 
+## Looks looked up: {{Appearance}}
+
+An image model draws a named figure from whatever it half remembers, so the same character can look different in every image. Put `{{Appearance}}` in an image, thumbnail or establishing prompt, for example a line `Looks: {{Appearance}}`, and Slopify looks the looks up instead.
+
+- Once the article exists, one call to the project's AI model **searches the web** for how the video's subject and every named character in the article look: build, face, clothing, colours and the marks that make them recognisable. Where depictions disagree, it describes the most iconic one.
+- With **Scenes from the article** on, each picture gets the looks of the figures its scene names (up to three besides the subject), so a scene of a place or of someone else isn't redrawn with the subject in it; a scene that names no one leaves the looks line out. A thumbnail always gets the subject's look. Without scenes, every picture gets the subject's look.
+- There is no switch: a prompt with `{{Appearance}}` turns it on, and prompts without it are drawn as they always were. `{{Appearance}}` is never a field to fill in.
+- It costs one more AI call per run, with web search. The pictures that use it wait for it.
+
 ## How long each image stays on screen
 
 **Seconds per image** (1 to 600, default 15) is in the **Video and style** row, beside **Cuts**, **Zoom** and **Motion**. When cuts follow the narration, it is the target length and a cut waits for a sentence end. When images run out, they start again. See [Play Video and Style](Play-Video-and-Style).

package/dist/tutorials/Play-Outputs.md

@@ -32,7 +32,7 @@ The Thumbnail part has its own **Source** switch:
 | Option | What it does | Default |
 | --- | --- | --- |
 | **Thumbnail prompt** | The Library thumbnail prompt, keywords filled in. **From prompt** sends it to the image model as it is; **Prompt by LLM** gives it to the text model with the title and article to write the image prompt. | None |
-| **Thumbnails** (**1** or **3**) | **3** makes two more thumbnails from the same prompt with different compositions, for YouTube's Test & compare. Each costs one more image, so 3 thumbnails are 3 image calls. | 1 |
+| **Thumbnails** (**1** or **3**) | **3** makes two more thumbnails from the same prompt with different compositions, for YouTube's Test & compare. Each costs one more image, so 3 thumbnails are 3 image calls. With a thumbnail prompt that has `{{Scene}}` and Images → **Scenes from the article** on, each also shows a different moment of the article (see [Play Images](Play-Images#scenes-from-the-article)). | 1 |
 
 ### Make three thumbnails for Test & compare