export type VideoModelType = 'image-to-video' | 'text-to-video';
export interface VideoModelSpec {
    id: string;
    name: string;
    type: VideoModelType;
    durations: string[];
    resolutions: string[];
    aspectRatios: string[];
    audio: boolean;
    audioConfigurable: boolean;
    audioInput: boolean;
    videoInput: boolean;
    /** Supports structured `elements` with @Element1/@Element2 prompt refs */
    supportsElements: boolean;
    /** Supports flat `reference_image_urls` array */
    supportsReferenceImages: boolean;
    /** Supports `scene_image_urls` for environment/style anchoring */
    supportsSceneImages: boolean;
    /** Supports `end_image_url` for targeted ending composition */
    supportsEndImage: boolean;
    /** Max duration in seconds */
    maxDurationSec: number;
    /**
     * Supports per-reference audio (Wan 2.7 R2V): each `elements[].audio_url`
     * can drive a different speaker's lip-sync inside a single render.
     */
    perReferenceAudio?: boolean;
    /**
     * Supports `reference_audio_urls` — voice-donor clips (up to 3, 2-15s each,
     * ≤15s aggregate, wav/mp3, ≤15MB per file) bound in-prompt as @Audio1,
     * @Audio2, … so a character's voice (timbre / accent / pacing) stays
     * consistent across shots. Must be paired with ≥1 reference image (Venice
     * rejects audio-only requests at validation). Distinct from `audioInput`
     * (lip-sync `audio_url`): these four R2V models accept reference audio but
     * do NOT set `audio_input: true` in GET /models. Confirmed live via
     * /video/quote (HTTP 200) 2026-07-23.
     */
    supportsReferenceAudio?: boolean;
    /**
     * Minimum allowed duration (seconds) for `audio_url` input.
     * Wan 2.7 rejects audio shorter than 3 seconds. Use the pre-flight
     * helper in `src/venice/audio-preflight.ts` to pad shorter clips.
     */
    minAudioInputSec?: number;
    /**
     * How much prompt the model wants.
     *
     * `'directorial'` (the default, and every family except MiniMax H3 Max) is
     * the house style: camera, blocking, location geography, aesthetic, and
     * reference role clauses all stated per shot, because these models render
     * what they are told and drift when they are not.
     *
     * `'simple'` models degrade with that treatment. They compose their own
     * coverage — framing, cutting, and beat rhythm — from one plain statement of
     * intent, and a wall of directorial clauses fights the shot the model would
     * have chosen. The prompt builder drops the heavy blocks for these (see
     * `buildVideoPrompt` / `buildMontagePrompt`) and keeps only what binds
     * identity and look.
     */
    promptStyle?: 'simple' | 'directorial';
    privacy: 'private' | 'anonymized';
    offline: boolean;
}
/**
 * True when a model wants a short, plain prompt rather than the full
 * directorial stack. Drives the lean branches in the mini-drama prompt
 * builder; unknown ids fall through to directorial (the safe default).
 */
export declare function modelWantsSimplePrompt(modelId: string): boolean;
/**
 * MiniMax H3 (all variants) image-to-video renders die server-side when the
 * START frame (`image_url`) contains a recognizable human face: Venice accepts
 * and bills the queue, then `/video/retrieve` 500s forever, with no
 * needs_consent handshake like Seedance. Loop chaining (which feeds each shot's
 * last frame into the next shot's i2v) and any panel-anchored i2v off a
 * face-bearing panel hit this. See AGENTS.md anti-pattern 31 (PR #25).
 *
 * Callers use this to avoid feeding a face-bearing start frame to these models
 * (degrade to t2v, skip face-continuity prompting, etc.). It's a per-model
 * capability so a future Venice fix — or a new i2v family — is a one-line change.
 */
export declare function i2vRejectsFaceStartFrame(modelId: string): boolean;
/**
 * Per-image-model positive-prompt length caps.
 *
 * Venice silently rejects requests with overly long positive prompts on
 * certain models (observed at ~1800-2200 chars on seedream-v5-lite; the
 * mini-drama character-reference builder used to emit 2400+ char prompts).
 *
 * Default cap is intentionally conservative (300 chars). Callers should
 * keep the most-important style cue + character anchor inline and move
 * everything else to negative_prompt.
 */
export declare const DEFAULT_MAX_POSITIVE_PROMPT_CHARS = 300;
export declare const MAX_POSITIVE_PROMPT_CHARS: Record<string, number>;
/**
 * Look up the positive-prompt cap for an image-generation model.
 * Falls back to DEFAULT_MAX_POSITIVE_PROMPT_CHARS when unspecified.
 */
export declare function getMaxPositivePromptChars(modelId: string): number;
export declare const VIDEO_MODELS: VideoModelSpec[];
export interface ImageModelSpec {
    id: string;
    name: string;
    type: 'generation' | 'edit' | 'upscale' | 'background-remove';
    offline: boolean;
}
export declare const IMAGE_GENERATION_MODELS: ImageModelSpec[];
export declare const MULTI_EDIT_MODELS: readonly ["qwen-edit", "qwen-image-2-edit", "qwen-image-2-pro-edit", "flux-2-max-edit", "gpt-image-1-5-edit", "gpt-image-2-edit", "grok-imagine-edit", "nano-banana-2-edit", "nano-banana-pro-edit", "seedream-v4-edit", "seedream-v5-lite-edit"];
export type MultiEditModelId = typeof MULTI_EDIT_MODELS[number];
export interface MusicModelSpec {
    id: string;
    name: string;
    type: 'music' | 'sound-effects' | 'tts';
    offline: boolean;
    /** Selectable voices (voice-enabled models). `default_voice` first if known. */
    voices?: string[];
    /** Default voice id. For seed-audio this is the sentinel "Describe in prompt". */
    defaultVoice?: string;
    /** `speed` param support + bounds. */
    supportsSpeed?: boolean;
    minSpeed?: number;
    maxSpeed?: number;
    defaultSpeed?: number;
    /** Lyrics / instrumental behaviour. */
    supportsLyrics?: boolean;
    lyricsRequired?: boolean;
    supportsForceInstrumental?: boolean;
    supportsLanguageCode?: boolean;
    /** Output containers the model can emit (`response_format` on retrieve). */
    supportedFormats?: string[];
    defaultFormat?: string;
    /** Prompt length bounds, in characters. */
    promptCharacterLimit?: number;
    minPromptLength?: number;
    /** Default generation length in seconds. */
    defaultDurationSec?: number;
    /** Per-second price in USD (for budgeting without a `/audio/quote` round-trip). */
    pricingPerSecondUsd?: number;
    /** One-line human summary. */
    description?: string;
}
export declare const MUSIC_MODELS: MusicModelSpec[];
export declare const TTS_MODELS: readonly ["tts-kokoro", "tts-qwen3-0-6b", "tts-qwen3-1-7b"];
export type TTSModelId = typeof TTS_MODELS[number];
export declare function getVideoModel(id: string): VideoModelSpec | undefined;
export declare function getMusicModel(id: string): MusicModelSpec | undefined;
export declare function listMusicModels(filter?: {
    type?: MusicModelSpec['type'];
}): MusicModelSpec[];
export declare function listVideoModels(filter?: {
    type?: VideoModelType;
    audio?: boolean;
    minDurationSec?: number;
    supportsElements?: boolean;
    supportsReferenceImages?: boolean;
    supportsEndImage?: boolean;
    imageToVideo?: boolean;
}): VideoModelSpec[];
/**
 * Check if a model supports a given duration string (e.g. "8s").
 * Falls back to checking max duration if the duration is within range.
 */
export declare function modelSupportsDuration(modelId: string, duration: string): boolean;
/**
 * For a given model, return the closest valid duration to the requested one.
 */
export declare function closestValidDuration(modelId: string, requestedSec: number): string | undefined;
/**
 * Venice's `bitrate_mode` for the Seedance family: `'high'` encodes at ~5-6x
 * the bitrate for sharper output and far fewer compression artifacts (larger
 * files), `'standard'` is the API default. It does NOT affect token price.
 */
export type BitrateMode = 'standard' | 'high';
/**
 * Default `bitrate_mode` applied to every Seedance 2.5 render. Seedance 2.5's
 * standard encode is visibly soft/blocky on detailed footage; `'high'` is the
 * dramatic-quality default and costs nothing extra.
 */
export declare const DEFAULT_SEEDANCE_25_BITRATE_MODE: BitrateMode;
/**
 * True for any Seedance 2.5 video model id — the registry ids
 * (`seedance-2-5-text-to-video`, `-image-to-video`, `-reference-to-video`) as
 * well as the `-basic` / `-i2v` id spellings some launch scripts use.
 */
export declare function isSeedance25VideoModel(modelId: string): boolean;
/**
 * Resolve the `bitrate_mode` to attach to a `/video/queue` body for a model.
 * An explicit override always wins; otherwise Seedance 2.5 defaults to `'high'`
 * and every other model returns `undefined` (the field is Seedance-only, so
 * callers should skip it rather than send it to models that reject it).
 */
export declare function resolveBitrateMode(modelId: string, override?: BitrateMode): BitrateMode | undefined;
/**
 * Build the model-specific parameters for a video queue request.
 * Handles resolution, aspect_ratio, and end_image_url based on model capabilities.
 */
export declare function buildModelParams(modelId: string, opts: {
    aspectRatio?: string;
    resolution?: string;
    endImageUrl?: string;
}): Record<string, unknown>;
//# sourceMappingURL=models.d.ts.map