vidspotai-shared 1.0.124 → 1.0.126

package/lib/services/aiGen/providers/bytedance/bytedance.service.js

@@ -4,6 +4,7 @@ var __importDefault = (this && this.__importDefault) || function (mod) {
 };
 Object.defineProperty(exports, "__esModule", { value: true });
 exports.ByteDanceService = void 0;
+exports.seedreamSize = seedreamSize;
 // bytedance.service.ts
 const baseAiGenProvider_service_1 = require("../baseAiGenProvider.service");
 const helpers_1 = require("../../helpers");
@@ -11,26 +12,57 @@ const aiModels_1 = require("../../../../globals/aiModels");
 const types_1 = require("../../../../globals/types");
 const firebase_1 = require("../../../../libs/firebase");
 const storage_1 = require("../../../../libs/storage");
+const image_model_1 = require("../../../../models/image.model");
 const helpers_2 = require("../../../../utils/helpers");
 const errors_1 = require("../../../../utils/errors");
 const logger_1 = require("../../../../utils/logger");
 const crypto_1 = __importDefault(require("crypto"));
+const SEEDREAM_URL = "https://ark.ap-southeast.bytepluses.com/api/v3/images/generations";
+/** Long-edge-equivalent of each `size` tier (tier² = target pixel count). */
+const SEEDREAM_TIER_EDGE = {
+    "1K": 1024,
+    "1.5K": 1536,
+    "2K": 2048,
+    "3K": 3072,
+    "4K": 4096,
+};
+/**
+ * Explicit "WxH" for a ratio at a size tier: tier² pixels in the requested
+ * shape, floored to multiples of 16. Every tier's pixel count sits inside the
+ * range the models that offer it accept (e.g. 5.0 lite: [2560×1440, 4096²]),
+ * and flooring never drops below it, so any ratio in [1/16, 16] is valid —
+ * including ones a model's aspect list doesn't name (the agent's 4:5).
+ */
+function seedreamSize(aspectRatio, tier) {
+    const edge = SEEDREAM_TIER_EDGE[tier];
+    const [w, h] = aspectRatio.split(":").map(Number);
+    if (!edge || !w || !h || w / h > 16 || h / w > 16)
+        return undefined;
+    const pixels = edge * edge;
+    const width = Math.floor(Math.sqrt((pixels * w) / h) / 16) * 16;
+    const height = Math.floor(Math.sqrt((pixels * h) / w) / 16) * 16;
+    return `${width}x${height}`;
+}
 class ByteDanceService extends baseAiGenProvider_service_1.BaseAiGenProviderService {
     constructor() {
         super(...arguments);
         this.baseUrl = "https://ark.ap-southeast.bytepluses.com/api/v3/contents/generations/tasks";
     }
-    // Bytedance Seedance only accepts ASCII printable characters and CJK (Chinese) characters.
-    // Strip everything else (emojis, en-dashes, special Unicode, etc.) and collapse whitespace.
-    sanitizePrompt(text) {
-        return text
-            .replace(/[^\x20-\x7E一-鿿㐀-䶿]/g, " ")
-            .replace(/\s+/g, " ")
-            .trim();
+    // Seedance 1.x only accepts ASCII printable characters and CJK (Chinese)
+    // characters: strip everything else (emojis, en-dashes, special Unicode,
+    // etc.) and collapse whitespace. The 2.x series takes prompts in 11 (2.5) /
+    // 7 (2.0) languages, including Arabic and Thai, so there we strip only emoji
+    // and control characters. Unsupported languages are translated upstream
+    // (supportedPromptLanguages → generate.controller).
+    sanitizePrompt(text, isV2) {
+        const stripped = isV2
+            ? text.replace(/[\p{Extended_Pictographic}\p{Cc}\u{FE0F}\u{200D}]/gu, " ")
+            : text.replace(/[^\x20-\x7E一-鿿㐀-䶿]/g, " ");
+        return stripped.replace(/\s+/g, " ").trim();
     }
     // Seedance 2.x (Dreamina — 2.0, 2.0-fast, 2.5) use a new multimodal content
     // schema with `role` tags on each non-text item, a `safety_identifier`, and
-    // no `service_tier`. Earlier Seedance models (1.x) reject the `role` field.
+    // no `service_tier`. 1.x takes `role` only for first+last-frame requests.
     // Matches any "bytedance-seedance-2." key so 2.5 (and future 2.x variants)
     // inherit the same request shape and video-input billing as 2.0.
     isSeedance2(modelKey) {
@@ -40,57 +72,91 @@ class ByteDanceService extends baseAiGenProvider_service_1.BaseAiGenProviderServ
         (0, helpers_1.validateParams)(params);
         const modelConfig = aiModels_1.aiModelConfigs[params.modelKey];
         const modelId = modelConfig.modelId;
-        const sanitizedPrompt = this.sanitizePrompt(params.prompt ?? "");
+        const isV2 = this.isSeedance2(params.modelKey);
+        const sanitizedPrompt = this.sanitizePrompt(params.prompt ?? "", isV2);
         if (!sanitizedPrompt) {
-            // PROMPT_INVALID (not an uncoded throw): sanitizePrompt strips everything
-            // outside ASCII+CJK, so an empty result means the prompt was written in a
-            // script Seedance cannot accept (Cyrillic, Arabic, Devanagari, Thai, an
-            // all-emoji prompt, …). That is deterministic user input — it can never
-            // succeed on retry. Without a code, `isNonPagingCode(undefined)` is false
-            // and the scene-failure handler logged it at `error`, which pages Slack
-            // (see errorPaging.ts). PROMPT_INVALID is in NON_PAGING_ERROR_CODES, so
-            // this now stays a warn while still failing the scene cleanly.
+            // PROMPT_INVALID (not an uncoded throw): an empty result means the prompt
+            // was written in a script this Seedance model cannot accept (1.x: anything
+            // outside ASCII+CJK; any model: an all-emoji prompt). That is
+            // deterministic user input — it can never succeed on retry. Without a
+            // code, `isNonPagingCode(undefined)` is false and the scene-failure
+            // handler logged it at `error`, which pages Slack (see errorPaging.ts).
+            // PROMPT_INVALID is in NON_PAGING_ERROR_CODES, so this stays a warn
+            // while still failing the scene cleanly.
             throw new errors_1.UserFacingError("Your prompt could not be processed. Please use English or Chinese text and try again.", errors_1.USER_FACING_ERROR_CODES.PROMPT_INVALID);
         }
-        const isV2 = this.isSeedance2(params.modelKey);
+        const fields = modelConfig.fields;
         const content = [
             {
                 type: "text",
                 text: sanitizedPrompt,
             },
         ];
-        // Image input. Seedance 2.0 expects role="first_frame" for I2V (or
-        // "reference_image" when combined with a video/audio reference). Earlier
-        // Seedance models reject `role` and treat any image as the first frame.
-        if (params.inputImageUrl) {
-            const item = {
-                type: "image_url",
-                image_url: { url: params.inputImageUrl },
-            };
-            if (isV2) {
-                // If a video reference is also present this is a multimodal reference,
-                // not first-frame I2V, so tag as reference_image.
-                item.role = params.inputVideoUrl ? "reference_image" : "first_frame";
-            }
-            content.push(item);
-        }
-        // Seedance 2.0 multimodal reference: video and/or audio inputs.
-        if (isV2) {
-            if (params.inputVideoUrl) {
-                content.push({
-                    type: "video_url",
-                    video_url: { url: params.inputVideoUrl },
-                    role: "reference_video",
+        // Omni reference-to-video (2.x only): reference images / videos / audio,
+        // each tagged with its role. Per docs.byteplus.com/en/docs/ModelArk/1520757
+        // first-frame, first+last-frame and omni-reference are mutually exclusive
+        // scenarios. `inputVideoUrl` / `inputAudioUrl` are the single-asset form
+        // the video routes send; the `reference*Urls` lists come from the agent.
+        const cap = (field, urls) => {
+            const max = fields[field]?.max;
+            return typeof max === "number" ? urls.slice(0, max) : urls;
+        };
+        const refVideos = isV2
+            ? cap("referenceVideoUrls", [
+                ...(params.inputVideoUrl ? [params.inputVideoUrl] : []),
+                ...(params.referenceVideoUrls ?? []),
+            ])
+            : [];
+        const refAudios = isV2
+            ? cap("referenceAudioUrls", [
+                ...(params.inputAudioUrl ? [params.inputAudioUrl] : []),
+                ...(params.referenceAudioUrls ?? []),
+            ])
+            : [];
+        const refImagesIn = isV2 ? (params.referenceImageUrls ?? []) : [];
+        const isOmni = refVideos.length + refAudios.length + refImagesIn.length > 0;
+        if (isOmni) {
+            // An input image alongside references can't be a first frame — it joins
+            // the reference images (the prompt can still say "@Image1 as the first
+            // frame", which is the documented workaround).
+            const refImages = cap("referenceImageUrls", [
+                ...(params.inputImageUrl ? [params.inputImageUrl] : []),
+                ...refImagesIn,
+            ]);
+            if (params.lastFrameImageUrl) {
+                logger_1.logger.warn("Seedance: last frame dropped — not combinable with omni references", {
+                    modelKey: params.modelKey,
                 });
             }
-            if (params.inputAudioUrl) {
-                content.push({
-                    type: "audio_url",
-                    audio_url: { url: params.inputAudioUrl },
-                    role: "reference_audio",
-                });
+            for (const url of refImages) {
+                content.push({ type: "image_url", image_url: { url }, role: "reference_image" });
+            }
+            for (const url of refVideos) {
+                content.push({ type: "video_url", video_url: { url }, role: "reference_video" });
+            }
+            for (const url of refAudios) {
+                content.push({ type: "audio_url", audio_url: { url }, role: "reference_audio" });
             }
         }
+        else if (params.inputImageUrl) {
+            // First frame, optionally with a last frame (2.x, 1.5 pro, 1.0 pro —
+            // the configs that declare lastFrameImageUrl; validateParams rejects it
+            // elsewhere). With two images the role is required on every model; with
+            // one it is optional, and we keep the 1.x request role-free as before.
+            const lastFrame = params.lastFrameImageUrl && "lastFrameImageUrl" in fields
+                ? params.lastFrameImageUrl
+                : undefined;
+            const tagRoles = isV2 || !!lastFrame;
+            content.push({
+                type: "image_url",
+                image_url: { url: params.inputImageUrl },
+                ...(tagRoles ? { role: "first_frame" } : {}),
+            });
+            if (lastFrame) {
+                content.push({ type: "image_url", image_url: { url: lastFrame }, role: "last_frame" });
+            }
+        }
+        const isFrameMode = !isOmni && !!params.inputImageUrl;
         const body = {
             model: modelId,
             content,
@@ -100,22 +166,31 @@ class ByteDanceService extends baseAiGenProvider_service_1.BaseAiGenProviderServ
             watermark: false,
             execution_expires_after: 172800, // 48h
         };
-        // For first-frame / first-last-frame generation (an input image present),
-        // ByteDance derives the output ratio from the image itself and rejects an
-        // explicit `ratio` (InvalidParameter.TaskTypeConstraint, prod 2026-08-11).
-        // Only pure text-to-video sends the requested aspect ratio.
-        if (!params.inputImageUrl) {
+        // For first-frame / first-last-frame generation, ByteDance derives the
+        // output ratio from the image itself and rejects an explicit `ratio`
+        // (InvalidParameter.TaskTypeConstraint, prod 2026-08-11). Text-to-video
+        // and omni reference-to-video take the requested aspect ratio.
+        if (!isFrameMode) {
             body.ratio = params.aspectRatio; // e.g. 16:9
         }
-        if (!this.isSeedance2(params.modelKey)) {
+        if (!isV2) {
             body.service_tier = "default";
         }
+        // `seed` and `camera_fixed` are 1.x-only per the docs; the configs that
+        // accept them declare the field. camera_fixed is unsupported with
+        // reference images (1.x has none, but keep the guard with the rule).
+        if (typeof params.seed === "number" && "seed" in fields) {
+            body.seed = params.seed;
+        }
+        if (params.cameraFixed && "cameraFixed" in fields && !isOmni) {
+            body.camera_fixed = true;
+        }
         // Byteplus KYC commitment K2 (deadline 2026-07-01): every Seedance 2.0
         // request must carry a stable per-user safety_identifier so Byteplus can
         // attribute downstream policy violations to a single end-user. We hash the
         // Firebase userId (sha256, 16 hex chars) — stable per user, but never
         // exposes the raw UID to the provider.
-        if (this.isSeedance2(params.modelKey) && params.userId) {
+        if (isV2 && params.userId) {
             body.safety_identifier = crypto_1.default
                 .createHash("sha256")
                 .update(params.userId)
@@ -294,6 +369,231 @@ class ByteDanceService extends baseAiGenProvider_service_1.BaseAiGenProviderServ
             status: types_1.EVideoSceneStatus.COMPLETED,
         };
     }
+    /**
+     * Seedream text-to-image / image-to-image. Three modes, each gated on the
+     * model declaring the field:
+     *   - default: one request per output image, run in parallel, so numImages
+     *     is exact.
+     *   - `imageSet` (5.0 lite / 4.x): one `sequential_image_generation: "auto"`
+     *     request returning a set of related images, at most numImages (the model
+     *     picks the count; the worker refunds the unused share).
+     *   - `layerDecomposition` (5.0 pro / flash): one input image → base image +
+     *     up to 16 transparent PNG layers, returned with their stacking order and
+     *     bounding boxes.
+     */
+    async generateImage(params) {
+        const modelConfig = aiModels_1.aiModelConfigs[params.modelKey];
+        if (!modelConfig?.modelId)
+            throw new Error(`Unknown image modelKey: ${params.modelKey}`);
+        const fields = modelConfig.fields;
+        const refs = [
+            ...(params.inputImageUrl ? [params.inputImageUrl] : []),
+            ...(params.inputImageUrls ?? []),
+        ];
+        const maxRefs = fields.inputImageUrls?.max ?? 0;
+        if (refs.length > maxRefs) {
+            logger_1.logger.warn("seedream: reference images capped", { modelKey: params.modelKey, given: refs.length, max: maxRefs });
+        }
+        const images = refs.slice(0, maxRefs);
+        const body = {
+            model: modelConfig.modelId,
+            prompt: params.prompt,
+            response_format: "url",
+            watermark: params.watermark ?? false,
+        };
+        const promptModes = fields.promptMode?.allowedValues;
+        if (params.promptMode && promptModes?.includes(params.promptMode)) {
+            body.optimize_prompt_options = { mode: params.promptMode };
+        }
+        const tiers = (fields.imageSize?.allowedValues ?? []);
+        const tier = params.imageSize && tiers.includes(params.imageSize)
+            ? params.imageSize
+            : fields.imageSize?.default;
+        if (params.layerDecomposition) {
+            if (!fields.layerDecomposition) {
+                throw new errors_1.UserFacingError("This model can't split an image into layers. Use Seedream 5.0 Pro or Flash.", errors_1.USER_FACING_ERROR_CODES.CAPABILITY_MISMATCH);
+            }
+            if (images.length !== 1) {
+                throw new errors_1.UserFacingError("Layer decomposition needs exactly one input image.", errors_1.USER_FACING_ERROR_CODES.CAPABILITY_MISMATCH);
+            }
+            // Tiers only (no WxH); without one, `auto` keeps the input's size.
+            body.layer_decomposition = true;
+            body.image = images[0];
+            body.size = params.imageSize && tiers.includes(params.imageSize) ? params.imageSize : "auto";
+            // output_format applies to the base image; layers are always png.
+            const formats = fields.outputFormat?.allowedValues;
+            if (params.outputFormat && formats?.includes(params.outputFormat))
+                body.output_format = params.outputFormat;
+            const { items, billed } = await this.requestSeedream(body, true);
+            const result = await this.storeSeedreamLayers(items.slice(0, image_model_1.SEEDREAM_MAX_LAYER_OUTPUTS));
+            // Pricing says "billed per successfully generated image" and the API
+            // docs' decomposition example reports usage.generated_images = base +
+            // layers. Settle on the provider's count either way, and log it so the
+            // first invoice can be checked against it.
+            logger_1.logger.info("seedream: layer decomposition usage", {
+                modelKey: params.modelKey,
+                outputs: items.length,
+                billed,
+            });
+            return billed !== undefined ? { ...result, billedOutputs: billed } : result;
+        }
+        if (params.dimensions && /^\d+x\d+$/.test(params.dimensions)) {
+            body.size = params.dimensions;
+        }
+        else if (tier) {
+            // No ratio (e.g. a plain edit) → the bare tier, which keeps the input's shape.
+            body.size = (params.aspectRatio && seedreamSize(params.aspectRatio, tier)) ?? tier;
+        }
+        if (images.length === 1)
+            body.image = images[0];
+        else if (images.length > 1)
+            body.image = images;
+        const formats = fields.outputFormat?.allowedValues;
+        const backgrounds = fields.background?.allowedValues;
+        // `background: transparent` works only as an edit of exactly one image
+        // that has alpha, and forces png.
+        const transparent = params.background === "transparent" && !!backgrounds?.includes("transparent") && images.length === 1;
+        if (transparent) {
+            body.background = "transparent";
+            body.output_format = "png";
+        }
+        else if (params.outputFormat && formats?.includes(params.outputFormat)) {
+            body.output_format = params.outputFormat;
+        }
+        const count = Math.max(1, Math.min(params.numImages ?? 1, 8));
+        let sources;
+        let billed;
+        if (params.imageSet && fields.imageSet && count > 1) {
+            // Inputs + outputs ≤ 15 per request.
+            body.sequential_image_generation = "auto";
+            body.sequential_image_generation_options = { max_images: Math.max(1, Math.min(count, 15 - images.length)) };
+            ({ items: sources, billed } = await this.requestSeedream(body, images.length > 0));
+        }
+        else {
+            const results = await Promise.all(Array.from({ length: count }, () => this.requestSeedream(body, images.length > 0)));
+            sources = results.flatMap((r) => r.items);
+            billed = results.every((r) => r.billed !== undefined)
+                ? results.reduce((sum, r) => sum + (r.billed ?? 0), 0)
+                : undefined;
+        }
+        const ts = Date.now();
+        const urls = [];
+        for (let i = 0; i < sources.length; i++) {
+            const item = sources[i];
+            let bytes;
+            let mime = item.output_format === "png" ? "image/png" : "image/jpeg";
+            if (item.b64_json)
+                bytes = Buffer.from(item.b64_json, "base64");
+            else if (item.url) {
+                // BytePlus URLs expire after 24h — copy the bytes now.
+                const r = await fetch(item.url);
+                if (!r.ok)
+                    throw new Error(`seedream image download ${r.status}`);
+                bytes = Buffer.from(await r.arrayBuffer());
+                mime = r.headers.get("content-type")?.split(";")[0] || mime;
+            }
+            else
+                continue;
+            const ext = mime === "image/png" ? "png" : mime === "image/webp" ? "webp" : "jpg";
+            const path = `images/bytedance/${ts}-${crypto_1.default.randomBytes(3).toString("hex")}-${i}.${ext}`;
+            urls.push(await storage_1.mediaStorage.put(path, bytes, mime));
+        }
+        if (!urls.length)
+            throw new Error("seedream returned no images");
+        const [width, height] = (sources[0]?.size ?? "").split("x").map(Number);
+        return {
+            imageUrl: urls[0],
+            imageUrls: urls.length > 1 ? urls : undefined,
+            ...(width && height ? { width, height } : {}),
+            ...(billed !== undefined ? { billedOutputs: billed } : {}),
+        };
+    }
+    /** Copy a layer decomposition's outputs into our storage, base image first. */
+    async storeSeedreamLayers(data) {
+        const ts = Date.now();
+        const sorted = [...data].sort((a, b) => (a.z_index ?? 0) - (b.z_index ?? 0));
+        const layers = [];
+        for (let i = 0; i < sorted.length; i++) {
+            const item = sorted[i];
+            let bytes;
+            let mime = item.output_format === "jpeg" ? "image/jpeg" : "image/png";
+            if (item.b64_json)
+                bytes = Buffer.from(item.b64_json, "base64");
+            else if (item.url) {
+                const r = await fetch(item.url);
+                if (!r.ok)
+                    throw new Error(`seedream layer download ${r.status}`);
+                bytes = Buffer.from(await r.arrayBuffer());
+                mime = r.headers.get("content-type")?.split(";")[0] || mime;
+            }
+            else
+                continue;
+            const ext = mime === "image/jpeg" ? "jpg" : mime === "image/webp" ? "webp" : "png";
+            const path = `images/bytedance/${ts}-${crypto_1.default.randomBytes(3).toString("hex")}-layer${i}.${ext}`;
+            const bbox = item.bounding_box?.normalized;
+            layers.push({
+                url: await storage_1.mediaStorage.put(path, bytes, mime),
+                zIndex: item.z_index ?? i,
+                ...(item.name ? { name: item.name } : {}),
+                ...(item.description ? { description: item.description } : {}),
+                ...(bbox?.length === 4 ? { bbox: [bbox[0], bbox[1], bbox[2], bbox[3]] } : {}),
+            });
+        }
+        if (!layers.length)
+            throw new Error("seedream layer decomposition returned no images");
+        const [width, height] = (sorted[0]?.size ?? "").split("x").map(Number);
+        return {
+            imageUrl: layers[0].url,
+            imageUrls: layers.map((l) => l.url),
+            layers,
+            ...(width && height ? { width, height } : {}),
+        };
+    }
+    /**
+     * One /images/generations call → its successful items (failed ones aren't
+     * billed) plus `usage.generated_images`, the count BytePlus bills.
+     */
+    async requestSeedream(body, isEdit) {
+        const resp = await fetch(SEEDREAM_URL, {
+            method: "POST",
+            headers: {
+                "Content-Type": "application/json",
+                Authorization: `Bearer ${process.env.BYTEDANCE_API_KEY}`,
+            },
+            body: JSON.stringify(body),
+        });
+        const text = await resp.text();
+        let json = {};
+        try {
+            json = JSON.parse(text);
+        }
+        catch {
+            /* non-JSON body — handled below */
+        }
+        const data = json.data ?? [];
+        const ok = data.filter((d) => d.url || d.b64_json);
+        if (resp.ok && ok.length) {
+            const billed = json.usage?.generated_images;
+            return { items: ok, billed: typeof billed === "number" ? billed : undefined };
+        }
+        const err = json.error ?? data.find((d) => d.error)?.error;
+        const code = err?.code ?? "";
+        const message = err?.message ?? text;
+        if (/SensitiveContentDetected/.test(code)) {
+            throw new errors_1.UserFacingError(code.startsWith("InputImage")
+                ? "Your input image contains sensitive content and cannot be processed. Please use a different image."
+                : "Your prompt or the generated image violates the content policy. Please modify your prompt and try again.", code.startsWith("InputImage")
+                ? errors_1.USER_FACING_ERROR_CODES.INPUT_IMAGE_REJECTED
+                : errors_1.USER_FACING_ERROR_CODES.CONTENT_POLICY_VIOLATION);
+        }
+        if (resp.status === 400 && isEdit && /image/i.test(`${code} ${json.error?.param ?? ""} ${message}`)) {
+            throw new errors_1.UserFacingError(message, errors_1.USER_FACING_ERROR_CODES.INPUT_IMAGE_REJECTED);
+        }
+        if (resp.status === 400) {
+            throw new errors_1.UserFacingError(message, errors_1.USER_FACING_ERROR_CODES.CAPABILITY_MISMATCH);
+        }
+        throw new Error(`seedream images failed (${resp.status}): ${code} ${message}`);
+    }
     getCreditUsed(params) {
         const modelConfig = aiModels_1.aiModelConfigs[params.modelKey];
         // Image-gen — flat per-image cost (Seedream). Must be handled before the
@@ -302,7 +602,24 @@ class ByteDanceService extends baseAiGenProvider_service_1.BaseAiGenProviderServ
         // image credit floor passed by the image controller.
         if (modelConfig?.cost?.fixed !== undefined &&
             modelConfig.type?.some((t) => t.endsWith("-to-image"))) {
-            return (0, helpers_2.getCreditsFromCost)(modelConfig.cost.fixed * Math.max(1, params.numImages ?? 1), !params.multiClip, params.creditFloor);
+            const cost = modelConfig.cost;
+            const tier = params.imageSize ?? modelConfig.fields.imageSize?.default;
+            const perImage = (tier ? cost.perImageSize?.[tier] : undefined) ?? cost.fixed;
+            // Layer decomposition bills each output (base + layers) and the count is
+            // the model's choice, so estimate the maximum; the image worker refunds
+            // the unused share, settling on the provider's usage.generated_images
+            // (see generateImage). API docs' decomposition example bills base + layers.
+            const images = params.layerDecomposition
+                ? image_model_1.SEEDREAM_MAX_LAYER_OUTPUTS
+                : Math.max(1, Math.min(params.numImages ?? 1, 8));
+            // 5.0 pro bills reference images after the first. Normal mode sends one
+            // request per output, so refs are billed once per image; an image set or
+            // a decomposition is a single request.
+            const extraRefs = Math.max(0, (params.numInputImages ?? 0) - 1);
+            const refCost = (cost.perExtraInputImage ?? 0) * extraRefs;
+            const singleRequest = params.layerDecomposition || params.imageSet;
+            const total = singleRequest ? perImage * images + refCost : (perImage + refCost) * images;
+            return (0, helpers_2.getCreditsFromCost)(total, !params.multiClip, params.creditFloor);
         }
         // Seedance 2.0 prices on (resolution × whether the request includes a
         // video input), not on (audio toggle). Pick the matching per-1K rate.

package/lib/services/aiGen/providers/google/google.service.d.ts

@@ -30,7 +30,23 @@ export declare class GoogleService extends BaseAiGenProviderService {
      */
     private withTransientRetry;
     generateVideo(params: VideoGenerationParams): Promise<VideoGenerationResult>;
+    /**
+     * Gemini Omni submit (Interactions API, `background: true`). Input mapping per
+     * the Omni guide: one image = image-to-video, two images = first/last-frame
+     * interpolation, several images = subject references. Omni has no
+     * negativePrompt / seed / duration-seconds params, so a negative prompt and
+     * the requested length are folded into the prompt text (the guide's own
+     * advice for negatives: "Do not do X").
+     */
+    private generateOmniVideo;
+    private checkOmniStatus;
     checkVideoStatus({ task, outputFilename, outputFilePath, }: VideoStatusParams): Promise<VideoStatusResult>;
+    /**
+     * Download a Files-API video (Veo output, or Gemini Omni `delivery:"uri"`
+     * output) with the key that produced it and persist it GCS-primary with the
+     * truncation guards below. Returns the scene's terminal status.
+     */
+    private persistFilesApiVideo;
     generateImage(params: ImageGenerationParams): Promise<ImageGenerationResult>;
     private _generateImage;
     /**

package/lib/services/aiGen/providers/google/google.service.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"google.service.d.ts","sourceRoot":"","sources":["../../../../../src/services/aiGen/providers/google/google.service.ts"],"names":[],"mappings":"AAkBA,OAAO,EAAE,wBAAwB,EAAE,MAAM,8BAA8B,CAAC;AACxE,OAAO,EACL,iBAAiB,EACjB,qBAAqB,EACrB,qBAAqB,EACrB,qBAAqB,EACrB,qBAAqB,EACrB,oBAAoB,EACpB,oBAAoB,EACpB,qBAAqB,EACrB,qBAAqB,EACrB,iBAAiB,EACjB,iBAAiB,EAClB,MAAM,UAAU,CAAC;AA+ElB,qBAAa,aAAc,SAAQ,wBAAwB;IAKzD,OAAO,CAAC,EAAE,CAAc;IACxB,OAAO,CAAC,OAAO,CAAgB;IAC/B,OAAO,CAAC,MAAM,CAAC,QAAQ,CAAC,kBAAkB,CAAK;;IAQ/C;;;;;;;;;OASG;IACG,YAAY,CAChB,MAAM,EAAE,oBAAoB,GAC3B,OAAO,CAAC,oBAAoB,CAAC;IAoChC;;;;;OAKG;IACH,OAAO,CAAC,eAAe;IAoCvB;;;;OAIG;YACW,kBAAkB;IAsC1B,aAAa,CACjB,MAAM,EAAE,qBAAqB,GAC5B,OAAO,CAAC,qBAAqB,CAAC;IAgU3B,gBAAgB,CAAC,EACrB,IAAI,EACJ,cAAc,EACd,cAAyB,GAC1B,EAAE,iBAAiB,GAAG,OAAO,CAAC,iBAAiB,CAAC;IAwK3C,aAAa,CACjB,MAAM,EAAE,qBAAqB,GAC5B,OAAO,CAAC,qBAAqB,CAAC;YAiBnB,cAAc;IAqK5B;;;;;;OAMG;IACH;;;;OAIG;IACG,aAAa,CACjB,MAAM,EAAE,qBAAqB,GAC5B,OAAO,CAAC,qBAAqB,CAAC;IAIjC,aAAa,CAAC,EAAE,QAAQ,EAAE,QAAY,EAAE,UAAmB,EAAE,SAAiB,EAAE,SAAa,EAAE,SAAS,EAAE,WAAW,EAAE,EAAE,iBAAiB,GAAG,MAAM;CA8BpJ"}
+{"version":3,"file":"google.service.d.ts","sourceRoot":"","sources":["../../../../../src/services/aiGen/providers/google/google.service.ts"],"names":[],"mappings":"AAqBA,OAAO,EAAE,wBAAwB,EAAE,MAAM,8BAA8B,CAAC;AACxE,OAAO,EACL,iBAAiB,EACjB,qBAAqB,EACrB,qBAAqB,EACrB,qBAAqB,EACrB,qBAAqB,EACrB,oBAAoB,EACpB,oBAAoB,EACpB,qBAAqB,EACrB,qBAAqB,EACrB,iBAAiB,EACjB,iBAAiB,EAClB,MAAM,UAAU,CAAC;AAuIlB,qBAAa,aAAc,SAAQ,wBAAwB;IAKzD,OAAO,CAAC,EAAE,CAAc;IACxB,OAAO,CAAC,OAAO,CAAgB;IAC/B,OAAO,CAAC,MAAM,CAAC,QAAQ,CAAC,kBAAkB,CAAK;;IAQ/C;;;;;;;;;OASG;IACG,YAAY,CAChB,MAAM,EAAE,oBAAoB,GAC3B,OAAO,CAAC,oBAAoB,CAAC;IAoChC;;;;;OAKG;IACH,OAAO,CAAC,eAAe;IAoCvB;;;;OAIG;YACW,kBAAkB;IAsC1B,aAAa,CACjB,MAAM,EAAE,qBAAqB,GAC5B,OAAO,CAAC,qBAAqB,CAAC;IAiUjC;;;;;;;OAOG;YACW,iBAAiB;YAoGjB,eAAe;IA4EvB,gBAAgB,CAAC,EACrB,IAAI,EACJ,cAAc,EACd,cAAyB,GAC1B,EAAE,iBAAiB,GAAG,OAAO,CAAC,iBAAiB,CAAC;IAyEjD;;;;OAIG;YACW,oBAAoB;IAkH5B,aAAa,CACjB,MAAM,EAAE,qBAAqB,GAC5B,OAAO,CAAC,qBAAqB,CAAC;YAiBnB,cAAc;IAkP5B;;;;;;OAMG;IACH;;;;OAIG;IACG,aAAa,CACjB,MAAM,EAAE,qBAAqB,GAC5B,OAAO,CAAC,qBAAqB,CAAC;IAIjC,aAAa,CAAC,EAAE,QAAQ,EAAE,QAAY,EAAE,UAAmB,EAAE,SAAiB,EAAE,SAAa,EAAE,SAAS,EAAE,WAAW,EAAE,EAAE,iBAAiB,GAAG,MAAM;CAoCpJ"}