STITCHER: Data Stitching Algorithms & Systems

• STITCHER is a family of methods that merge fragmented data from modalities including images, sensors, and latent representations using unified inpainting and alignment techniques.
• The framework leverages single-pass algorithms, weighted masks, and low-dimensional mappings to achieve computational efficiency, robust output coherence, and adaptability across varied inputs.
• Applications span from image stitching and spectral cube integration to robot trajectory planning, semantic segmentation, and forensic data aggregation, demonstrating versatile use cases.

STITCHER refers to a family of algorithms, frameworks, and systems designed to join, align, or merge data fragments from diverse modalities—images, time series, physical sensors, latent representations, or even software artifacts—into unified outputs. In both computational imaging and broader domains, “stitching” mechanisms enable the integration of partially overlapping or fragmented data streams, correcting for inconsistencies, filling gaps, and enforcing global coherence. The technical realization of STITCHER varies by field: from unified inpainting approaches for image fusion to graph-based motion primitive composition in robotics and scalable latent-map alignment in end-to-end AI systems. Across these instances, STITCHER solutions are characterized by a deliberate focus on computational efficiency, interpretability, and adaptability to heterogeneity in the input data.

1. Image Stitching: Unified Inpainting and Beyond

Traditional image stitching pipelines comprise three interconnected stages: registration (alignment), fusion (blending overlapped regions), and rectangling (hole-filling and boundary normalization). Historically, each stage required separate optimization, often resulting in error propagation and intricate parameter dependencies. The Simple and Robust Stitcher (SRStitcher) introduces a paradigm shift by reframing both fusion and rectangling as a single, guided inpainting problem. The fusion stage modifies pixels only within an overlap mask using a blending or reconstruction operator, while rectangling fills masked boundary regions via hole-filling operators. SRStitcher unifies these via:

$$I_{\rm out}(x,y) = \mathrm{Denoise}(I_{\rm in}(x,y), M(x,y), P(x,y)),$$

where $I_{\rm in}$ is a coarse-stitched panorama, $M$ is a composite mask indicating inpainting regions, and $P$ is a continuous weighted mask guiding pixel-level diffusion strength. This pipeline leverages pre-trained diffusion models (e.g., Stable-Diffusion-2-inpainting) in a single inference pass, requiring neither new training nor fine-tuning. Weighted masks—constructed by content, seam, and distance transforms—control the blend/inpaint intensity across seam and hole regions, ensuring localized modifications and interpretable edit maps. On the UDIS-D benchmark, SRStitcher achieves top hyperIQA and CLIPIQA scores, with expert and lay evaluations preferring it over previous methods in >70% of cases (Xie et al., 2024).

Variants such as RDIStitcher further generalize the approach by combining a larger modification area and stronger inpainting intensity based on diffusion U-Nets and reference-driven pseudo-tasks, yielding superior seam elimination and zero-shot robustness, as evaluated by MLLM-based quality metrics (Xie et al., 2024). Both approaches indicate the trend toward self-contained, end-to-end differentiable image stitchers that subsume traditional pipeline steps and allow flexible reparameterization or prompt adaptation.

2. Spectral, Mapping, and Physical Data Stitching

The necessity for stitching extends into astronomy, crystallography, and sensor networks, demanding rigorous handling of overlaps, intensity mismatches, and spatially variant resolutions.

In JWST/MIRI spectral cube stitching (Canin et al., 17 Nov 2025), each IFU sub-cube partially overlaps spectrally and spatially with adjacent cubes, but features distinct intensity scales and sampling grids. The Haute Couture algorithm performs:

• Global intensity homogenization by solving a small system for scale coefficients over overlaps, ensuring flux continuity.
• Matrix completion via nonnegative matrix factorization (NMF), treating sub-cubes as blocks of a larger incomplete matrix and imputing missing entries by low-rank factorization with β-divergence minimization. This resolves both overlap discontinuities and spatial sparsity in a single optimization, maintaining maximal spatial resolution and minimizing propagation error.

Similarly, RSstitcher merges 2D X-ray diffraction frames into Wide Range Reciprocal Space Maps, employing precise geometric coordinate transformations and reciprocal-space intensity merging. The algorithm applies law-of-cosines–based angle calculations and merges overlapping pixels by maximal intensity, providing artifact-free reciprocal maps for both symmetric and side-inclination GID geometries (Wang et al., 11 Nov 2025). In underwater video mosaicking, CORSTitch applies Fourier-domain correlation, phase-only correlation, and precise GNSS-time alignment to robustly merge geospatial mosaics with sub-pixel accuracy (Maypa et al., 1 May 2025).

3. Trajectory, Policy, and Latent Representation Stitchers

STITCHER frameworks in robotics and autonomous AI optimize the composition of local or modular policies, motion segments, or latent encodings for global consistency, dynamic feasibility, and rapid planning.

The real-time motion primitive STITCHER (Levy et al., 2024, Levy et al., 16 Oct 2025) for trajectory planning addresses the intractability of nonlinear programming in high-dimensional, constraint-rich spaces. It proceeds in three stages:

1. Geometric path search: A* or Dijkstra finds coarse, obstacle-free waypoints.
2. Velocity graph/DP: Double-integrator, minimum-time primitives generate a pruned, structured graph over discrete velocities; dynamic programming yields cost-to-go heuristics.
3. A* over full-state motion primitives: Higher-order closed-form primitives (triple integrator, jerk-limited) connect waypoints under exact constraints, using prior DP results as admissible heuristics. Each primitive is checked via uniform sampling for collision and state/actuator compliance.

All data structures are finite and statically allocated, providing millisecond-scale planning guarantees across complex environments, with bounding optimality relative to sampling resolution. This outperforms optimization-based planners (e.g., MIQP solvers) by several orders of magnitude in compute time while maintaining hard constraint satisfaction.

For end-to-end perception-action pipelines in autonomous driving, a distinct STITCHER module aligns the outputs of updated perception encoders to frozen downstream driving policies (Li et al., 19 Jun 2026). Low-complexity linear or convolutional “stitchers” learn an affine or 1×1 convolutional mapping in latent space, trained on paired anchor samples to ensure downstream policy compatibility. Empirically, convolutional stitchers recover >91% of the original driving score after major sensor or domain shifts and require three orders less adaptation time than full retraining. These results support the low-dimensional manifold hypothesis: as long as both encoders remain task-aligned, a low-complexity transformation suffices for functional equivalence.

4. Stitching in Representation Learning, Segmentation, and LLMs

In the context of hierarchical representation learning, language modeling, and dense prediction, STITCHER modules enforce overlap-invariance and global context after localized, windowed, or fragmented processing.

Zonkey's hierarchical diffusion LLM introduces a Stitcher that merges overlapping, soft-masked latent segments, each derived from differentiable tokenization (Segment Splitter) with probabilistic BOS boundaries. The Stitcher estimates soft alignment offsets based on content similarity and existence drops, blends overlapping positions by existence-weighted averaging, and applies a local Transformer refinement with probabilistic attention. This process enforces that more than one segment's “opinion” dominates at overlap, yielding near-identity latent codes in redundant regions and facilitating end-to-end gradient flow. Ablations demonstrate that omission of the Stitcher leads to significantly higher overlap error and degraded global coherence (Rozental, 29 Jan 2026).

In open-vocabulary semantic segmentation at scale, OV-Stitcher reconstructs global self-attention by “stitching” the final encoder block's Q/K/V representations of overlapping sub-crops, enabling cross-crop global reasoning even under sliding-window input constraints. The “Stitch Attention” mechanism transforms fragmented feature maps into consistent representations, eliminating patch-boundary artifacts and delivering +2–3 mIoU improvements over previous training-free baselines (Moon et al., 9 Apr 2026).

5. STITCHER Approaches in Object Detection, Audio, and Forensics

Adaptive data-stitching is instrumental for scale-robust training and robust inference.

In object detection, Dynamic Scale Training (DST/STITCHER) uses runtime loss statistics on small-object instances to dynamically switch between regular and “collaged” multi-scale input batches. This non-parametric controller eliminates the need for a fixed multi-scale schedule, substantially accelerates convergence, and delivers +1.7–+2.7 AP improvements on COCO detection without architectural modification or inference overhead (Chen et al., 2020).

In audio deepfake detection, “neural stitching” refers to replacement of the dense penultimate nonlinearity with a direct softmax connection to the embedding during inference. This empirically suppresses overfitting on seen generator artifacts, driving large relative reductions in error rates on adversarially generated fakes, and highlights the role of linear downstream heads for out-of-distribution robustness (Yan et al., 2022).

In digital forensic correlation on IoT, STITCHER denotes a multi-stage evidence aggregation pipeline combining classification, extraction, and correlation algorithms to integrate firmware, network, and process artifacts. Automated set-intersection and difference routines facilitate rapid anomaly and linkage discovery, providing reproducible forensic reports and substantially reducing time-to-solve in controlled user studies (Tok et al., 2020).

6. Implementation Practices, Failure Modes, and Future Directions

Across STITCHER systems, key technical practices include:

• Explicit mask, weight, or boundary constructs for controlling modification location/intensity.
• Single-pass or batchwise inference/optimization for scalability.
• Priors or constraints enforcing continuity, invariance, or minimal duplication across overlaps.
• Hyperparameterization of kernel/scaling/overlap and ablation-driven parameter selection.
• Quantitative and mixed human-in-the-loop evaluation for output consistency and interpretability.

Failure cases often arise from excessive gap regions, color or domain mismatches exceeding the capabilities of a single inpainting or affine alignment, or excessive fragmentation/duplication overwhelming the downstream fusion logic. Future directions emphasize: extension to multi-layer or multi-modal architectures, learnable or kernelized alignment/stitching, external auxiliary loss incorporation for stability, scaling to larger or more irregularly sampled input spaces, and increased integration with standard data frameworks in each application domain.

STITCHER methods thus constitute a foundational family of strategies for merging, aligning, and reconstructing coherent outputs from modular or partially redundant fragments, enabling both localized precision and global consistency across a wide spectrum of computational and experimental sciences.