framefields 2.9.2 → 2.10.0-canary.0

package/dist/index-DQa7mhzq.d.mts

@@ -0,0 +1,330 @@
+import { A as VisionCompute, L as VisionSession, M as VisionModelDescriptor, N as VisionModelStore, P as VisionModelStoreOptions, R as VisionTensorInput, T as SessionProvider, j as VisionImageInput, z as VisionTensorOutputs } from "./index-WwuuX8ZJ.mjs";
+
+//#region ../vision-sam/dist/index.d.mts
+//#region src/license.d.ts
+declare const SAM_LICENSE_TERMS_URL = "https://github.com/facebookresearch/sam3/blob/main/LICENSE";
+interface SamLicenseOptions {
+  /** Explicitly agree to Meta's SAM License terms to download or run SAM 3 models. */
+  readonly acceptLicense?: boolean;
+}
+/**
+ * Ensures that the user has explicitly accepted Meta's SAM License terms before
+ * downloading SAM 3 weights or performing inference. SAM 2.1's weights are Apache-2.0 and
+ * need no acceptance.
+ */
+declare function assertSamLicenseAccepted(context?: string, options?: SamLicenseOptions): void;
+//#endregion
+//#region src/models/registry.d.ts
+/**
+ * SAM model registry. Every file is pinned to an immutable Hugging Face revision with its
+ * exact size and SHA-256; the shared `VisionModelStore` verifies both after download.
+ * Nothing here touches the network: the descriptors are registered into the vision
+ * registry on first tracker creation so `@framefields/vision` never lists SAM weights.
+ */
+/** SAM 2.1 checkpoints with a pinned video export. */
+type Sam2Model = "tiny" | "small";
+type Sam2Graph = "vision_encoder" | "mask_decoder" | "memory_encoder" | "memory_attention" | "pointer_tpos";
+type Sam2ModelKey = `sam2.1-${Sam2Model}-${Sam2Graph | "constants"}`;
+/**
+ * Diffusion Studio's fp16 video exports of Meta SAM 2.1 (Apache-2.0). Tiny is built for
+ * speed at 512² (32×32 features); Small runs SAM 2's native 1024² (64×64), resolving small
+ * objects, thin structures and identity after cuts better at roughly 4× the encoder and
+ * 16× the memory-attention work.
+ */
+declare const SAM2_EXPORTS: Readonly<Record<Sam2Model, {
+  repo: string;
+  revision: string;
+  size: number;
+  name: string;
+}>>;
+/** Weight formats published for the SAM 3 tracker's image encoder. */
+type Sam3Precision = "fp16" | "q4f16" | "q4";
+/**
+ * The encoder precision each backend runs fastest. Measured on an M1 Pro (one 1008² frame):
+ * WebGPU fp16 3.65 s, q4f16 4.43 s, q4 6.63 s, int8 140 s (its ops fall back to the CPU);
+ * fp16 is also the closest to full precision. On WebGPU, 4-bit weights are dequantized
+ * before every matmul, so they only save memory. CUDA runs fp16 on tensor cores. CPUs have
+ * no fast fp16 path, so they get fp32-activation q4.
+ */
+declare const SAM3_ENCODER: Readonly<Record<VisionCompute, Sam3Precision>>;
+type Sam3ModelKey = `sam3-vision_encoder-${Sam3Precision}` | `sam3-vision_encoder-${Sam3Precision}-data` | "sam3-mask_decoder" | "sam3-memory_encoder" | "sam3-memory_attention" | "sam3-support" | "sam3-support-data" | "sam3-text-vision_encoder" | "sam3-text-vision_encoder-data" | "sam3-text-encoder" | "sam3-text-decoder" | "sam3-tokenizer" | "sam3-tokenizer-config";
+type SamModelKey = Sam2ModelKey | Sam3ModelKey;
+declare const SAM_LICENSE = "Meta SAM License";
+/** File names of the encoder at `precision`; the graph refers to its weights by `data`. */
+declare function sam3EncoderFiles(precision: Sam3Precision): {
+  graph: string;
+  data: string;
+};
+declare const SAM_MODELS: readonly VisionModelDescriptor[];
+/** Registers the SAM descriptors into the shared vision registry (idempotent). */
+declare function ensureSamModelsRegistered(): void;
+//#endregion
+//#region src/sam2/memory.d.ts
+/** CPU-owned memory bank; neural inference uses the configured vision execution provider. */
+interface Sam2Memory {
+  readonly frame: number;
+  readonly tokens: Float32Array;
+  readonly positions: Float32Array;
+  readonly pointer: Float32Array;
+}
+declare class Sam2MemoryBank {
+  private readonly temporal;
+  private seed?;
+  private recent;
+  constructor(temporal: readonly (readonly number[])[]);
+  remember(memory: Sam2Memory, prompted: boolean): void;
+  assemble(frame: number, frameCount?: number): {
+    memory: Float32Array;
+    positions: Float32Array;
+    distances: Float32Array;
+  };
+}
+/** NCHW features to the sequence-major layout consumed by memory attention. */
+declare function sam2Tokens(data: Float32Array): Float32Array;
+//#endregion
+//#region src/sam2/types.d.ts
+/** Normalized source-image coordinates. Labels 2/3 form a box's opposite corners. */
+interface Sam2Point {
+  readonly x: number;
+  readonly y: number;
+  readonly label: 0 | 1 | 2 | 3;
+}
+/** Frame is on the source analysis clock, before trim or time remapping. */
+interface Sam2Prompt {
+  readonly frame: number;
+  readonly points: readonly Sam2Point[];
+}
+interface Sam2Options {
+  readonly prompts: readonly Sam2Prompt[];
+  /** Metadata for object and class signals; SAM does not classify objects. */
+  readonly category?: string;
+  /**
+   * Checkpoint: `"tiny"` (default; 512², ~0.3 s/frame on an M1 Pro) or `"small"` (SAM 2's
+   * native 1024²; finer edges, small objects and thin structures, steadier identity after
+   * cuts, at several times the cost).
+   */
+  readonly model?: Sam2Model;
+}
+interface Sam2FrameResult {
+  readonly mask: Uint8Array;
+  readonly width: number;
+  readonly height: number;
+  readonly score: number;
+  readonly iou: number;
+}
+interface Sam2FrameTracker {
+  process(image: VisionImageInput, frame: number, points?: readonly Sam2Point[]): Promise<Sam2FrameResult>;
+  release(): void;
+}
+declare function validateSam2Prompts(prompts: readonly Sam2Prompt[]): void;
+//#endregion
+//#region src/sam2/tracker.d.ts
+/** The pinned export revision of `model`: part of ClipVision's persistent cache identity. */
+declare function sam2Revision(model?: Sam2Model): string;
+interface Sam2TrackerOptions extends VisionModelStoreOptions {
+  readonly provider?: SessionProvider;
+  readonly store?: VisionModelStore;
+  /** Source sequence length; used to normalize object-pointer time positions for short videos. */
+  readonly frameCount?: number;
+  /** Checkpoint: `"tiny"` (default, 512², fast) or `"small"` (1024², finer detail). */
+  readonly model?: Sam2Model;
+}
+/** Complete pinned SAM 2.1 video graph pipeline (Tiny or Small), one explicitly selected object. */
+declare class Sam2Tracker {
+  private readonly constants;
+  private readonly frameCount;
+  private readonly sessions;
+  private readonly bank;
+  /** Model input side; image features are `size / 16` square, one memory token per cell. */
+  private readonly size;
+  private readonly tokens;
+  private lastFrame;
+  private released;
+  private constructor();
+  static create(options?: Sam2TrackerOptions): Promise<Sam2Tracker>;
+  private run;
+  process(image: VisionImageInput, frame: number, points?: readonly Sam2Point[]): Promise<Sam2FrameResult>;
+  private condition;
+  private remember;
+  release(): void;
+}
+/** Bilinear resize of square logits before thresholding, matching SAM's source-image mask contract. */
+declare function decodeSam2Mask(logits: Float32Array, width: number, height: number, score: number, iou: number): Sam2FrameResult;
+//#endregion
+//#region src/sam3/graphs.d.ts
+declare function sam3SupportGraphs(reference: Uint8Array, external: Uint8Array): {
+  densePrompt: Uint8Array;
+  pointerPosition: Uint8Array;
+  temporal: Float32Array;
+  noMemory: Float32Array;
+};
+/** Replace the no-mask embedding by the actual dense prompt head's output. */
+declare function sam3DenseDecoder(modelBytes: Uint8Array): Uint8Array;
+//#endregion
+//#region src/sam3/memory.d.ts
+/** Dynamic memory lengths: never duplicate unavailable early-frame memories. */
+declare class Sam3MemoryBank {
+  private readonly temporal;
+  private seed?;
+  private recent;
+  constructor(temporal: Float32Array);
+  remember(memory: Sam2Memory, prompted: boolean): void;
+  assemble(frame: number, frameCount: number): {
+    spatial: Float32Array;
+    positions: Float32Array;
+    pointers: Float32Array;
+    encoding: Float32Array;
+    pointerCount: number;
+  };
+}
+/** SAM 3 PositionEmbeddingSine: y then x, normalized pixel centers at grid indices 1..72. */
+declare function sam3Position(): Float32Array;
+//#endregion
+//#region src/sam3/text-selector.d.ts
+declare function sam3Floats(outputs: VisionTensorOutputs, name: string): Float32Array;
+declare function sam3Run(session: VisionSession, inputs: readonly VisionTensorInput[]): Promise<VisionTensorOutputs>;
+/** Semantic model is separate from the visual tracker and loads only for text prompts. */
+declare class Sam3TextSelector {
+  private readonly tokenizer;
+  private readonly sessions;
+  private encoded?;
+  private constructor();
+  static create(store: VisionModelStore, provider: SessionProvider): Promise<Sam3TextSelector>;
+  /**
+   * Mask logits (288²) of the instances matching `text` with sigmoid score of at least
+   * `threshold`, best first and at most `max`. Queries that repeat an accepted instance
+   * (mask IoU above 0.5) are skipped. Throws when nothing matches unless `required` is false.
+   */
+  select(pixels: Float32Array, text: string, threshold: number, max?: number, required?: boolean): Promise<Float32Array[]>;
+  /** Text features for a prompt; the last one is kept, since re-detection repeats it every few frames. */
+  private encodeText;
+  release(): void;
+}
+//#endregion
+//#region src/sam3/types.d.ts
+interface Sam3Prompt {
+  /** Source analysis frame, before trimming or time remapping. */
+  readonly frame: number;
+  /** Normalized source coordinates; labels 2/3 define a box. */
+  readonly points?: readonly Sam2Point[];
+  /** Select matching instances (up to `maxObjects`, best first), then propagate their masks. */
+  readonly text?: string;
+  /**
+   * Track ID (1-based) a points prompt seeds or corrects; defaults to 1. The next unused
+   * ID adds an object. Text prompts re-select every object, so they take no ID.
+   */
+  readonly object?: number;
+}
+interface Sam3Options extends SamLicenseOptions {
+  /** Prompt list across frames. Optional if shorthand `text` is supplied. */
+  readonly prompts?: readonly Sam3Prompt[];
+  /** Convenience shorthand to select by text at frame 0. */
+  readonly text?: string;
+  /** Signal category metadata; defaults to "object". */
+  readonly category?: string;
+  /** Minimum sigmoid score for a text-selected instance. Default 0.5. */
+  readonly textThreshold?: number;
+  /** Most instances a text prompt selects, as track IDs 1…N by score. Default 1, or every match with `merge`. */
+  readonly maxObjects?: number;
+  /**
+   * Select every instance matching the text as **one** object: one mask and one memory, so
+   * tracking costs the same as a single object however many match (e.g. a whole crowd).
+   */
+  readonly merge?: boolean;
+  /**
+   * Every this many frames after a text prompt, run that text again and add the instances no
+   * tracked object covers yet (people who walk in, turn up from behind someone, or were lost):
+   * into the one object with `merge`, otherwise as new track IDs up to `maxObjects`. Each
+   * detection pass costs about one text prompt. Default 0: select once, then only track.
+   */
+  readonly redetect?: number;
+  /** Image-encoder weights; defaults to the fastest for the backend in use (`SAM3_ENCODER`). */
+  readonly precision?: Sam3Precision;
+}
+declare function validateSam3Prompts(options: Sam3Options): void;
+/** One tracked SAM 3 object at a frame; an empty mask means it is absent there. */
+interface Sam3ObjectResult {
+  /** Track ID, 1-based and stable until a text prompt re-selects every object. */
+  readonly id: number;
+  /** Source-sized binary mask (0 or 255). */
+  readonly mask: Uint8Array;
+  /** Object-presence logit; positive means present. */
+  readonly score: number;
+  readonly iou: number;
+}
+interface Sam3FrameResult {
+  readonly width: number;
+  readonly height: number;
+  /** Every tracked object, ordered by track ID. */
+  readonly objects: readonly Sam3ObjectResult[];
+}
+//#endregion
+//#region src/sam3/tracker.d.ts
+/** Every graph and tokenizer revision, and the encoder precision: ClipVision's cache identity. */
+declare function sam3Revision(precision: Sam3Precision): string;
+/** The encoder precision a tracker on `provider` uses: the override, else the backend's best. */
+declare function sam3Precision(provider?: SessionProvider, override?: Sam3Precision): Promise<Sam3Precision>;
+interface Sam3TrackerOptions extends VisionModelStoreOptions {
+  readonly provider?: SessionProvider;
+  readonly store?: VisionModelStore;
+  readonly frameCount?: number;
+  readonly textThreshold?: number;
+  /** Most instances a text prompt selects. Default 1, or every match with `merge`. */
+  readonly maxObjects?: number;
+  /** Track every instance a text prompt selects as one object (one mask, one memory). */
+  readonly merge?: boolean;
+  /** Re-run the last text prompt every this many frames and add what no object covers (see `Sam3Options`). */
+  readonly redetect?: number;
+  /** Image-encoder weights; defaults to the fastest for the provider's backend (`SAM3_ENCODER`). */
+  readonly precision?: Sam3Precision;
+  /** Explicitly agree to Meta's SAM License terms to download or run SAM 3's weights. */
+  readonly acceptLicense?: boolean;
+}
+/**
+ * Explicitly selected SAM 3 instances with full neural memory propagation on the provider.
+ * The image encoder runs once per frame; each object keeps its own memory bank.
+ */
+declare class Sam3Tracker {
+  private readonly store;
+  private readonly provider;
+  private readonly options;
+  private readonly sessions;
+  private encoder?;
+  private readonly position;
+  private readonly banks;
+  private temporal?;
+  private noMemory?;
+  private selector?;
+  /** The last text prompt and its frame, which re-detection repeats. */
+  private text?;
+  private lastFrame;
+  private released;
+  private constructor();
+  static create(options?: Sam3TrackerOptions): Promise<Sam3Tracker>;
+  private run;
+  process(image: VisionImageInput, frame: number, prompt?: Omit<Sam3Prompt, "frame">): Promise<Sam3FrameResult>;
+  private track;
+  /**
+   * On a `redetect` frame, runs the last text prompt again and seeds what no tracked object
+   * covers: merged into the one object (keeping what it tracks), or as new objects.
+   */
+  private redetect;
+  private textSelector;
+  /** A 288² concept mask as a seed that outputs (and remembers) that mask. */
+  private textSeed;
+  private densePrompt;
+  /** Objects a prompt (re)seeds, by index. Text re-selects every object; points address one. */
+  private seeds;
+  private newBank;
+  private decode;
+  private initialFeatures;
+  private condition;
+  private remember;
+  release(): void;
+}
+/** Resize logits before thresholding; source pixels are the public mask coordinate system. */
+declare function sam3Mask(logits: Float32Array, width: number, height: number, score: number): Uint8Array;
+//#endregion
+//#endregion
+export { assertSamLicenseAccepted as A, sam3Precision as B, Sam3Precision as C, Sam3TrackerOptions as D, Sam3Tracker as E, sam3DenseDecoder as F, validateSam3Prompts as G, sam3Run as H, sam3EncoderFiles as I, sam3Floats as L, ensureSamModelsRegistered as M, sam2Revision as N, SamLicenseOptions as O, sam2Tokens as P, sam3Mask as R, Sam3Options as S, Sam3TextSelector as T, sam3SupportGraphs as U, sam3Revision as V, validateSam2Prompts as W, Sam2TrackerOptions as _, SAM_MODELS as a, Sam3ModelKey as b, Sam2Graph as c, Sam2Model as d, Sam2ModelKey as f, Sam2Tracker as g, Sam2Prompt as h, SAM_LICENSE_TERMS_URL as i, decodeSam2Mask as j, SamModelKey as k, Sam2Memory as l, Sam2Point as m, SAM3_ENCODER as n, Sam2FrameResult as o, Sam2Options as p, SAM_LICENSE as r, Sam2FrameTracker as s, SAM2_EXPORTS as t, Sam2MemoryBank as u, Sam3FrameResult as v, Sam3Prompt as w, Sam3ObjectResult as x, Sam3MemoryBank as y, sam3Position as z };
+//# sourceMappingURL=index-DQa7mhzq.d.mts.map

package/dist/index-DQa7mhzq.d.mts.map

@@ -0,0 +1 @@
+{"version":3,"file":"index-DQa7mhzq.d.mts","names":["SessionProvider","VisionCompute","VisionImageInput","VisionModelDescriptor","VisionModelStore","VisionModelStoreOptions","VisionSession","VisionTensorInput","VisionTensorOutputs","SAM_LICENSE_TERMS_URL","SamLicenseOptions","assertSamLicenseAccepted","Sam2Model","Sam2Graph","Sam2ModelKey","SAM2_EXPORTS","Record","Readonly","Sam3Precision","SAM3_ENCODER","Sam3ModelKey","SamModelKey","SAM_LICENSE","sam3EncoderFiles","SAM_MODELS","ensureSamModelsRegistered","Sam2Memory","Float32Array","Sam2MemoryBank","sam2Tokens","Sam2Point","Sam2Prompt","Sam2Options","Sam2FrameResult","Uint8Array","Sam2FrameTracker","Promise","validateSam2Prompts","sam2Revision","Sam2TrackerOptions","Sam2Tracker","decodeSam2Mask","sam3SupportGraphs","sam3DenseDecoder","Sam3MemoryBank","sam3Position","sam3Floats","sam3Run","Sam3TextSelector","Sam3Prompt","Sam3Options","validateSam3Prompts","Sam3ObjectResult","Sam3FrameResult","sam3Revision","sam3Precision","Sam3TrackerOptions","Sam3Tracker","Omit","sam3Mask"],"sources":["../../vision-sam/dist/index.d.mts"],"sourcesContent":["import { SessionProvider, VisionCompute, VisionImageInput, VisionModelDescriptor, VisionModelStore, VisionModelStoreOptions, VisionSession, VisionTensorInput, VisionTensorOutputs } from \"@framefields/vision\";\n\n//#region src/license.d.ts\ndeclare const SAM_LICENSE_TERMS_URL = \"https://github.com/facebookresearch/sam3/blob/main/LICENSE\";\ninterface SamLicenseOptions {\n  /** Explicitly agree to Meta's SAM License terms to download or run SAM 3 models. */\n  readonly acceptLicense?: boolean;\n}\n/**\n * Ensures that the user has explicitly accepted Meta's SAM License terms before\n * downloading SAM 3 weights or performing inference. SAM 2.1's weights are Apache-2.0 and\n * need no acceptance.\n */\ndeclare function assertSamLicenseAccepted(context?: string, options?: SamLicenseOptions): void;\n//#endregion\n//#region src/models/registry.d.ts\n/**\n * SAM model registry. Every file is pinned to an immutable Hugging Face revision with its\n * exact size and SHA-256; the shared `VisionModelStore` verifies both after download.\n * Nothing here touches the network: the descriptors are registered into the vision\n * registry on first tracker creation so `@framefields/vision` never lists SAM weights.\n */\n/** SAM 2.1 checkpoints with a pinned video export. */\ntype Sam2Model = \"tiny\" | \"small\";\ntype Sam2Graph = \"vision_encoder\" | \"mask_decoder\" | \"memory_encoder\" | \"memory_attention\" | \"pointer_tpos\";\ntype Sam2ModelKey = `sam2.1-${Sam2Model}-${Sam2Graph | \"constants\"}`;\n/**\n * Diffusion Studio's fp16 video exports of Meta SAM 2.1 (Apache-2.0). Tiny is built for\n * speed at 512² (32×32 features); Small runs SAM 2's native 1024² (64×64), resolving small\n * objects, thin structures and identity after cuts better at roughly 4× the encoder and\n * 16× the memory-attention work.\n */\ndeclare const SAM2_EXPORTS: Readonly<Record<Sam2Model, {\n  repo: string;\n  revision: string;\n  size: number;\n  name: string;\n}>>;\n/** Weight formats published for the SAM 3 tracker's image encoder. */\ntype Sam3Precision = \"fp16\" | \"q4f16\" | \"q4\";\n/**\n * The encoder precision each backend runs fastest. Measured on an M1 Pro (one 1008² frame):\n * WebGPU fp16 3.65 s, q4f16 4.43 s, q4 6.63 s, int8 140 s (its ops fall back to the CPU);\n * fp16 is also the closest to full precision. On WebGPU, 4-bit weights are dequantized\n * before every matmul, so they only save memory. CUDA runs fp16 on tensor cores. CPUs have\n * no fast fp16 path, so they get fp32-activation q4.\n */\ndeclare const SAM3_ENCODER: Readonly<Record<VisionCompute, Sam3Precision>>;\ntype Sam3ModelKey = `sam3-vision_encoder-${Sam3Precision}` | `sam3-vision_encoder-${Sam3Precision}-data` | \"sam3-mask_decoder\" | \"sam3-memory_encoder\" | \"sam3-memory_attention\" | \"sam3-support\" | \"sam3-support-data\" | \"sam3-text-vision_encoder\" | \"sam3-text-vision_encoder-data\" | \"sam3-text-encoder\" | \"sam3-text-decoder\" | \"sam3-tokenizer\" | \"sam3-tokenizer-config\";\ntype SamModelKey = Sam2ModelKey | Sam3ModelKey;\ndeclare const SAM_LICENSE = \"Meta SAM License\";\n/** File names of the encoder at `precision`; the graph refers to its weights by `data`. */\ndeclare function sam3EncoderFiles(precision: Sam3Precision): {\n  graph: string;\n  data: string;\n};\ndeclare const SAM_MODELS: readonly VisionModelDescriptor[];\n/** Registers the SAM descriptors into the shared vision registry (idempotent). */\ndeclare function ensureSamModelsRegistered(): void;\n//#endregion\n//#region src/sam2/memory.d.ts\n/** CPU-owned memory bank; neural inference uses the configured vision execution provider. */\ninterface Sam2Memory {\n  readonly frame: number;\n  readonly tokens: Float32Array;\n  readonly positions: Float32Array;\n  readonly pointer: Float32Array;\n}\ndeclare class Sam2MemoryBank {\n  private readonly temporal;\n  private seed?;\n  private recent;\n  constructor(temporal: readonly (readonly number[])[]);\n  remember(memory: Sam2Memory, prompted: boolean): void;\n  assemble(frame: number, frameCount?: number): {\n    memory: Float32Array;\n    positions: Float32Array;\n    distances: Float32Array;\n  };\n}\n/** NCHW features to the sequence-major layout consumed by memory attention. */\ndeclare function sam2Tokens(data: Float32Array): Float32Array;\n//#endregion\n//#region src/sam2/types.d.ts\n/** Normalized source-image coordinates. Labels 2/3 form a box's opposite corners. */\ninterface Sam2Point {\n  readonly x: number;\n  readonly y: number;\n  readonly label: 0 | 1 | 2 | 3;\n}\n/** Frame is on the source analysis clock, before trim or time remapping. */\ninterface Sam2Prompt {\n  readonly frame: number;\n  readonly points: readonly Sam2Point[];\n}\ninterface Sam2Options {\n  readonly prompts: readonly Sam2Prompt[];\n  /** Metadata for object and class signals; SAM does not classify objects. */\n  readonly category?: string;\n  /**\n   * Checkpoint: `\"tiny\"` (default; 512², ~0.3 s/frame on an M1 Pro) or `\"small\"` (SAM 2's\n   * native 1024²; finer edges, small objects and thin structures, steadier identity after\n   * cuts, at several times the cost).\n   */\n  readonly model?: Sam2Model;\n}\ninterface Sam2FrameResult {\n  readonly mask: Uint8Array;\n  readonly width: number;\n  readonly height: number;\n  readonly score: number;\n  readonly iou: number;\n}\ninterface Sam2FrameTracker {\n  process(image: VisionImageInput, frame: number, points?: readonly Sam2Point[]): Promise<Sam2FrameResult>;\n  release(): void;\n}\ndeclare function validateSam2Prompts(prompts: readonly Sam2Prompt[]): void;\n//#endregion\n//#region src/sam2/tracker.d.ts\n/** The pinned export revision of `model`: part of ClipVision's persistent cache identity. */\ndeclare function sam2Revision(model?: Sam2Model): string;\ninterface Sam2TrackerOptions extends VisionModelStoreOptions {\n  readonly provider?: SessionProvider;\n  readonly store?: VisionModelStore;\n  /** Source sequence length; used to normalize object-pointer time positions for short videos. */\n  readonly frameCount?: number;\n  /** Checkpoint: `\"tiny\"` (default, 512², fast) or `\"small\"` (1024², finer detail). */\n  readonly model?: Sam2Model;\n}\n/** Complete pinned SAM 2.1 video graph pipeline (Tiny or Small), one explicitly selected object. */\ndeclare class Sam2Tracker {\n  private readonly constants;\n  private readonly frameCount;\n  private readonly sessions;\n  private readonly bank;\n  /** Model input side; image features are `size / 16` square, one memory token per cell. */\n  private readonly size;\n  private readonly tokens;\n  private lastFrame;\n  private released;\n  private constructor();\n  static create(options?: Sam2TrackerOptions): Promise<Sam2Tracker>;\n  private run;\n  process(image: VisionImageInput, frame: number, points?: readonly Sam2Point[]): Promise<Sam2FrameResult>;\n  private condition;\n  private remember;\n  release(): void;\n}\n/** Bilinear resize of square logits before thresholding, matching SAM's source-image mask contract. */\ndeclare function decodeSam2Mask(logits: Float32Array, width: number, height: number, score: number, iou: number): Sam2FrameResult;\n//#endregion\n//#region src/sam3/graphs.d.ts\ndeclare function sam3SupportGraphs(reference: Uint8Array, external: Uint8Array): {\n  densePrompt: Uint8Array;\n  pointerPosition: Uint8Array;\n  temporal: Float32Array;\n  noMemory: Float32Array;\n};\n/** Replace the no-mask embedding by the actual dense prompt head's output. */\ndeclare function sam3DenseDecoder(modelBytes: Uint8Array): Uint8Array;\n//#endregion\n//#region src/sam3/memory.d.ts\n/** Dynamic memory lengths: never duplicate unavailable early-frame memories. */\ndeclare class Sam3MemoryBank {\n  private readonly temporal;\n  private seed?;\n  private recent;\n  constructor(temporal: Float32Array);\n  remember(memory: Sam2Memory, prompted: boolean): void;\n  assemble(frame: number, frameCount: number): {\n    spatial: Float32Array;\n    positions: Float32Array;\n    pointers: Float32Array;\n    encoding: Float32Array;\n    pointerCount: number;\n  };\n}\n/** SAM 3 PositionEmbeddingSine: y then x, normalized pixel centers at grid indices 1..72. */\ndeclare function sam3Position(): Float32Array;\n//#endregion\n//#region src/sam3/text-selector.d.ts\ndeclare function sam3Floats(outputs: VisionTensorOutputs, name: string): Float32Array;\ndeclare function sam3Run(session: VisionSession, inputs: readonly VisionTensorInput[]): Promise<VisionTensorOutputs>;\n/** Semantic model is separate from the visual tracker and loads only for text prompts. */\ndeclare class Sam3TextSelector {\n  private readonly tokenizer;\n  private readonly sessions;\n  private encoded?;\n  private constructor();\n  static create(store: VisionModelStore, provider: SessionProvider): Promise<Sam3TextSelector>;\n  /**\n   * Mask logits (288²) of the instances matching `text` with sigmoid score of at least\n   * `threshold`, best first and at most `max`. Queries that repeat an accepted instance\n   * (mask IoU above 0.5) are skipped. Throws when nothing matches unless `required` is false.\n   */\n  select(pixels: Float32Array, text: string, threshold: number, max?: number, required?: boolean): Promise<Float32Array[]>;\n  /** Text features for a prompt; the last one is kept, since re-detection repeats it every few frames. */\n  private encodeText;\n  release(): void;\n}\n//#endregion\n//#region src/sam3/types.d.ts\ninterface Sam3Prompt {\n  /** Source analysis frame, before trimming or time remapping. */\n  readonly frame: number;\n  /** Normalized source coordinates; labels 2/3 define a box. */\n  readonly points?: readonly Sam2Point[];\n  /** Select matching instances (up to `maxObjects`, best first), then propagate their masks. */\n  readonly text?: string;\n  /**\n   * Track ID (1-based) a points prompt seeds or corrects; defaults to 1. The next unused\n   * ID adds an object. Text prompts re-select every object, so they take no ID.\n   */\n  readonly object?: number;\n}\ninterface Sam3Options extends SamLicenseOptions {\n  /** Prompt list across frames. Optional if shorthand `text` is supplied. */\n  readonly prompts?: readonly Sam3Prompt[];\n  /** Convenience shorthand to select by text at frame 0. */\n  readonly text?: string;\n  /** Signal category metadata; defaults to \"object\". */\n  readonly category?: string;\n  /** Minimum sigmoid score for a text-selected instance. Default 0.5. */\n  readonly textThreshold?: number;\n  /** Most instances a text prompt selects, as track IDs 1…N by score. Default 1, or every match with `merge`. */\n  readonly maxObjects?: number;\n  /**\n   * Select every instance matching the text as **one** object: one mask and one memory, so\n   * tracking costs the same as a single object however many match (e.g. a whole crowd).\n   */\n  readonly merge?: boolean;\n  /**\n   * Every this many frames after a text prompt, run that text again and add the instances no\n   * tracked object covers yet (people who walk in, turn up from behind someone, or were lost):\n   * into the one object with `merge`, otherwise as new track IDs up to `maxObjects`. Each\n   * detection pass costs about one text prompt. Default 0: select once, then only track.\n   */\n  readonly redetect?: number;\n  /** Image-encoder weights; defaults to the fastest for the backend in use (`SAM3_ENCODER`). */\n  readonly precision?: Sam3Precision;\n}\ndeclare function validateSam3Prompts(options: Sam3Options): void;\n/** One tracked SAM 3 object at a frame; an empty mask means it is absent there. */\ninterface Sam3ObjectResult {\n  /** Track ID, 1-based and stable until a text prompt re-selects every object. */\n  readonly id: number;\n  /** Source-sized binary mask (0 or 255). */\n  readonly mask: Uint8Array;\n  /** Object-presence logit; positive means present. */\n  readonly score: number;\n  readonly iou: number;\n}\ninterface Sam3FrameResult {\n  readonly width: number;\n  readonly height: number;\n  /** Every tracked object, ordered by track ID. */\n  readonly objects: readonly Sam3ObjectResult[];\n}\n//#endregion\n//#region src/sam3/tracker.d.ts\n/** Every graph and tokenizer revision, and the encoder precision: ClipVision's cache identity. */\ndeclare function sam3Revision(precision: Sam3Precision): string;\n/** The encoder precision a tracker on `provider` uses: the override, else the backend's best. */\ndeclare function sam3Precision(provider?: SessionProvider, override?: Sam3Precision): Promise<Sam3Precision>;\ninterface Sam3TrackerOptions extends VisionModelStoreOptions {\n  readonly provider?: SessionProvider;\n  readonly store?: VisionModelStore;\n  readonly frameCount?: number;\n  readonly textThreshold?: number;\n  /** Most instances a text prompt selects. Default 1, or every match with `merge`. */\n  readonly maxObjects?: number;\n  /** Track every instance a text prompt selects as one object (one mask, one memory). */\n  readonly merge?: boolean;\n  /** Re-run the last text prompt every this many frames and add what no object covers (see `Sam3Options`). */\n  readonly redetect?: number;\n  /** Image-encoder weights; defaults to the fastest for the provider's backend (`SAM3_ENCODER`). */\n  readonly precision?: Sam3Precision;\n  /** Explicitly agree to Meta's SAM License terms to download or run SAM 3's weights. */\n  readonly acceptLicense?: boolean;\n}\n/**\n * Explicitly selected SAM 3 instances with full neural memory propagation on the provider.\n * The image encoder runs once per frame; each object keeps its own memory bank.\n */\ndeclare class Sam3Tracker {\n  private readonly store;\n  private readonly provider;\n  private readonly options;\n  private readonly sessions;\n  private encoder?;\n  private readonly position;\n  private readonly banks;\n  private temporal?;\n  private noMemory?;\n  private selector?;\n  /** The last text prompt and its frame, which re-detection repeats. */\n  private text?;\n  private lastFrame;\n  private released;\n  private constructor();\n  static create(options?: Sam3TrackerOptions): Promise<Sam3Tracker>;\n  private run;\n  process(image: VisionImageInput, frame: number, prompt?: Omit<Sam3Prompt, \"frame\">): Promise<Sam3FrameResult>;\n  private track;\n  /**\n   * On a `redetect` frame, runs the last text prompt again and seeds what no tracked object\n   * covers: merged into the one object (keeping what it tracks), or as new objects.\n   */\n  private redetect;\n  private textSelector;\n  /** A 288² concept mask as a seed that outputs (and remembers) that mask. */\n  private textSeed;\n  private densePrompt;\n  /** Objects a prompt (re)seeds, by index. Text re-selects every object; points address one. */\n  private seeds;\n  private newBank;\n  private decode;\n  private initialFeatures;\n  private condition;\n  private remember;\n  release(): void;\n}\n/** Resize logits before thresholding; source pixels are the public mask coordinate system. */\ndeclare function sam3Mask(logits: Float32Array, width: number, height: number, score: number): Uint8Array;\n//#endregion\nexport { SAM2_EXPORTS, SAM3_ENCODER, SAM_LICENSE, SAM_LICENSE_TERMS_URL, SAM_MODELS, Sam2FrameResult, Sam2FrameTracker, Sam2Graph, Sam2Memory, Sam2MemoryBank, Sam2Model, Sam2ModelKey, Sam2Options, Sam2Point, Sam2Prompt, Sam2Tracker, Sam2TrackerOptions, Sam3FrameResult, Sam3MemoryBank, Sam3ModelKey, Sam3ObjectResult, Sam3Options, Sam3Precision, Sam3Prompt, Sam3TextSelector, Sam3Tracker, Sam3TrackerOptions, SamLicenseOptions, SamModelKey, assertSamLicenseAccepted, decodeSam2Mask, ensureSamModelsRegistered, sam2Revision, sam2Tokens, sam3DenseDecoder, sam3EncoderFiles, sam3Floats, sam3Mask, sam3Position, sam3Precision, sam3Revision, sam3Run, sam3SupportGraphs, validateSam2Prompts, validateSam3Prompts };\n//# sourceMappingURL=index.d.mts.map"],"mappings":";;;;AAAgN,cAGlMS,qBAAAA,GAAqB,4DAAA;AAAA,UACzBC,iBAAAA,CAAiB;EAAA;EAS4D,SAUlFE,aAAS,CAAA,EAAA,OAAA;AAAA;AACA;AACsC;;;;iBAZnCD,wBAAAA,CAmBmB,OAAA,CAAA,EAAA,MAAA,EAAA,OAAA,CAAA,EAnBkCD,iBAmBlC,CAAA,EAAA,IAAA;AAAA;AAOlB;;;;;;AAQkB;AAC6D;AACnD,KA1BzCE,SAAAA,GA2BSU,MAAW,GAAA,OAAA;AAAA,KA1BpBT,SAAAA,GA4BYU,gBAAgB,GAAYL,cAAa,GAAA,gBAAA,GAAA,kBAAA,GAAA,cAAA;AAAA,KA3BrDJ,YAAAA,GA+BqD,UA/B5BF,SA+BKT,IA/BQU,SA+Ba,GAAA,WAAA,EAAA;AAAA;AAEd;;;;;AAQV,cAlClBE,YAoCc,EApCAE,QAoCA,CApCSD,MAoCT,CApCgBJ,SAoChB,EAAA;EAKTc,IAAAA,EAAAA,MAAAA;EAEPC,QAAAA,EAAAA,MAAAA;EACGA,IAAAA,EAAAA,MAAAA;EACAA,IAAAA,EAAAA,MAAAA;CAAY,CAAA,CAAA;AAAA;AAIkC,KA1CxDT,aAAAA,GA8Cc,MAAA,GAAA,OAAA,GAAA,IAAA;AAAA;AAQkB;AAWT;AAGD;;;;cA5DbC,YAmEoEiB,EAnEtDnB,QAmEsDmB,CAnE7CpB,MAmE6CoB,CAnEtCnC,aAmEsCmC,EAnEvBlB,aAmEuBkB,CAAAA,CAAAA;KAlE7EhB,YAAAA,GAkEoF,uBAlE9CF,aAkE8C,EAAA,GAAA,uBAlELA,aAkEK,OAAA,GAAA,mBAAA,GAAA,qBAAA,GAAA,uBAAA,GAAA,cAAA,GAAA,mBAAA,GAAA,0BAAA,GAAA,+BAAA,GAAA,mBAAA,GAAA,mBAAA,GAAA,gBAAA,GAAA,uBAAA;AAAA,KAjEpFG,WAAAA,GAAcP,YAoEiB,GApEFM,YAoEqBW;AAAU,cAnEnDT,WAAAA,GAuEe,kBAAkB;AAAA;iBArE9BC,gBAAAA,CAuEKvB,SAAAA,EAvEuBkB,aAuEvBlB,CAAAA,EAAAA;EACHI,KAAAA,EAAAA,MAAAA;EAIAQ,IAAAA,EAAAA,MAAAA;CANkBP;cAlEvBmB,UAkE8C,EAAA,SAlEzBrB,qBAkEyB,EAAA;AAAA;iBAhE3CsB,yBAAAA,CAAAA,CAoFSc,EAAAA,IAAAA;;;;UAhFhBb,UAAAA,CAkF0DI;EAAsBG,SAAAA,KAAAA,EAAAA,MAAAA;EAARG,SAAAA,MAAAA,EAhF/DT,YAgF+DS;EAAO,SAAA,SAAA,EA/EnET,YA+EmE;EAAA,SAMxEc,OAAAA,EApFGd,YAoFW;AAAkG;cAlFnHC,cAAAA,CAqFgCM;EAAsBA,iBAAAA,QAAAA;EACrDA,QAAAA,IAAAA;EACIA,QAAAA,MAAAA;EACPP,WAAAA,CAAAA,QAAAA,EAAAA,SAAAA,CAAAA,SAAAA,MAAAA,EAAAA,CAAAA,EAAAA;EACAA,QAAAA,CAAAA,MAAAA,EApFOD,UAoFPC,EAAAA,QAAAA,EAAAA,OAAAA,CAAAA,EAAAA,IAAAA;EAAY,QAAA,CAAA,KAAA,EAAA,MAAA,EAAA,UAAA,CAAA,EAAA,MAAA,CAAA,EAAA;IAGPgB,MAAAA,EArFLhB,YAqFqB;IAInBiB,SAAAA,EAxFCjB,YAwFa;IAIJA,SAAAA,EA3FTA,YA2FSA;EACLD,CAAAA;;;iBAxFFG,UAAAA,CA4FHF,IAAAA,EA5FoBA,YA4FpBA,CAAAA,EA5FmCA,YA4FnCA;;;AACY;AAKmB,UA9FnCG,SAAAA,CAiGiB;EAA0D,SACpEiB,CAAAA,EAAO,MAAA;EAAUzC,SAAAA,CAAAA,EAAAA,MAAAA;EAAgCC,SAAAA,KAAAA,EAAAA,CAAAA,GAAAA,CAAAA,GAAAA,CAAAA,GAAAA,CAAAA;;;UA5FxDwB,UAAAA,CA4FqF;EAAA,SAEjFiB,KAAAA,EAAAA,MAAgB;EAKP5C,SAAAA,MAAAA,EAAAA,SAjGK0B,SAiGL1B,EAAAA;;UA/Fb4B,WAAAA,CA+FmEgB;EAARZ,SAAAA,OAAAA,EAAAA,SA9FxCL,UA8FwCK,EAAAA;EAMpDT;EAA0FA,SAAAA,QAAAA,CAAAA,EAAAA,MAAAA;EAARS;;AAAO;AAWpE;;EAiCflB,SAAAA,KAAAA,CAAAA,EAxIJN,SAwIIM;;UAtIbe,eAAAA,CA8GqC;EAAA,SA0B9BkB,IAAAA,EAvIAjB,UAuImB;EAAqB,SAE/CkB,KAAAA,EAAAA,MAAgB;EAIC,SAKjBC,MAAAA,EAAAA,MAAe;EAIoB,SAK5BC,KAAAA,EAAAA,MAAY;EAAyB,SAErCC,GAAAA,EAAAA,MAAa;;UAvJpBpB,gBAAAA,CAuJ4DjB;EAAwBA,OAAAA,CAAAA,KAAAA,EAtJ7EhB,gBAsJ6EgB,EAAAA,KAAAA,EAAAA,MAAAA,EAAAA,MAAAA,CAAAA,EAAAA,SAtJ1BY,SAsJ0BZ,EAAAA,CAAAA,EAtJZkB,OAsJYlB,CAtJJe,eAsJIf,CAAAA;EAARkB,OAAAA,EAAAA,EAAAA,IAAAA;;AAAO,iBAnJ5EC,mBAAAA,CAoJW,OAAA,EAAA,SApJ2BN,UAoJ3B,EAAA,CAAA,EAAA,IAAA;;;;iBAhJXO,YAAAA,CAgJoBjC,KAAAA,CAAAA,EAhJCO,SAgJDP,CAAAA,EAAAA,MAAAA;UA/I3BkC,kBAAAA,SAA2BlC,uBA+IuB,CAAA;EAAA,SAoB9CoD,QAAW,CAAA,EAlKHzD,eAkKG;EAgBCwD,SAAAA,KAAAA,CAAAA,EAjLPpD,gBAiLOoD;EAA6BC;EAARrB,SAAAA,UAAAA,CAAAA,EAAAA,MAAAA;EAE9BlC;EAA+C+C,SAAAA,KAAAA,CAAAA,EA/K7CrC,SA+K6CqC;;;cA5KlDT,WAAAA,CA4KyEJ;EAAO,iBAAA,SAAA;EAAA,iBAqBrE,UAAST;;;;;;;;;0BAtLRY,qBAAqBH,QAAQI;;iBAEtCtC,mDAAmD4B,cAAcM,QAAQH;;;;;;iBAMzEQ,cAAAA,SAAuBd,0EAA0EM;;;iBAGjGS,iBAAAA,YAA6BR,sBAAsBA;eACrDA;mBACIA;YACPP;YACAA;;;iBAGKgB,gBAAAA,aAA6BT,aAAaA;;;;cAI7CU,cAAAA;;;;wBAIUjB;mBACLD;;aAENC;eACEA;cACDA;cACAA;;;;;iBAKGkB,YAAAA,CAAAA,GAAgBlB;;;iBAGhBmB,UAAAA,UAAoBtC,oCAAoCmB;iBACxDoB,OAAAA,UAAiBzC,gCAAgCC,sBAAsB6B,QAAQ5B;;cAElFwC,gBAAAA;;;;;uBAKS5C,4BAA4BJ,kBAAkBoC,QAAQY;;;;;;iBAM5DrB,kFAAkFS,QAAQT;;;;;;;UAOjGsB,UAAAA;;;;6BAImBnB;;;;;;;;;UASnBoB,WAAAA,SAAoBxC;;8BAEAuC;;;;;;;;;;;;;;;;;;;;;;uBAsBP/B;;iBAENiC,mBAAAA,UAA6BD;;UAEpCE,gBAAAA;;;;iBAIOlB;;;;;UAKPmB,eAAAA;;;;6BAImBD;;;;;iBAKZE,YAAAA,YAAwBpC;;iBAExBqC,aAAAA,YAAyBvD,4BAA4BkB,gBAAgBkB,QAAQlB;UACpFsC,kBAAAA,SAA2BnD;sBACfL;mBACHI;;;;;;;;;;uBAUIc;;;;;;;;cAQTuC,WAAAA;;;;;;;;;;;;;;;;0BAgBYD,qBAAqBpB,QAAQqB;;iBAEtCvD,0CAA0CwD,KAAKT,uBAAuBb,QAAQiB;;;;;;;;;;;;;;;;;;;;;iBAqB9EM,QAAAA,SAAiBhC,6DAA6DO"}