/**
 * @fileoverview GaussianDensify — adaptive density control for the native 3DGS trainer.
 *
 * The original 3DGS paper's Algorithm 1: gaussians whose accumulated screen-space (2D mean) gradient
 * exceeds a threshold are "under-reconstructed" and get densified — small ones are CLONED (a jittered
 * copy is added), large ones are SPLIT (replaced by two smaller ones sampled from their distribution).
 * Transparent gaussians (opacity below a floor) are PRUNED. This is how a sparse init grows into a
 * dense scene — the mechanism the 2026-06-20 /critic review correctly named as missing
 * ("fixed-cardinality trainer"). Now built, gated, and verified.
 *
 * Deterministic: all randomness comes from a seeded RNG the caller passes, so densification is
 * reproducible (and the count is assertable in tests).
 */
import { type Gaussian3D } from './GaussianTrainer3D';
/** Per-gaussian densification statistics accumulated over an interval by the runner. */
export interface DensifyStats {
    /** Average ||∂L/∂mean2d|| (screen-space, px) per gaussian over the interval. */
    avgGrad2d: Float64Array;
}
export interface DensifyOpts {
    /** Densify gaussians whose avg 2D-mean gradient exceeds this (px). */
    gradThreshold: number;
    /** Prune gaussians with opacity below this. */
    opacityPrune: number;
    /** Clone if max world scale <= this, otherwise split. */
    scaleThreshold: number;
    /** Split shrink factor (paper: φ = 1.6). */
    splitFactor?: number;
    /** Hard cap on gaussian count (stop densifying when reached). */
    maxGaussians: number;
}
export interface DensifyResult {
    gaussians: Gaussian3D;
    /** origin[newIdx] = source oldIdx for survivors, -1 for freshly created (clone/split) gaussians.
     *  Lets the caller carry Adam moments for survivors and zero them for new ones. */
    origin: Int32Array;
}
/**
 * Apply pruning + densification. Returns a new (resized) Gaussian3D and the origin map. Gaussian
 * count grows (clones/splits) and shrinks (prune); a fixed-cardinality run is what you get by simply
 * not calling this.
 */
export declare function densifyAndPrune(g: Gaussian3D, stats: DensifyStats, opts: DensifyOpts, rng: () => number): DensifyResult;
/** Small seeded LCG so densification (and its tests) are deterministic. */
export declare function seededRng(seed: number): () => number;
//# sourceMappingURL=GaussianDensify.d.ts.map