Stitched Evaluation Protocol

• Stitched Evaluation Protocol is a structured evaluation design that assesses quality by integrating multiple aligned inputs rather than isolated global scores.
• It systematically assembles partial evidence from various views, network fragments, and phases to capture local distortions and artifacts missed by conventional metrics.
• The protocol has been applied across domains such as image quality, depth estimation, model stitching, and agent evaluation to enhance sensitivity and reproducibility.

A stitched evaluation protocol is an evaluation procedure in which the quantity of interest is assessed not from an isolated object or a single global score, but from a deliberately stitched construction: multiple views are aligned into a common domain, multiple network fragments are connected through a restricted adaptor, multiple phases are composed into a standardized pipeline, or multiple local outputs are assembled and then judged in a way that exposes artifacts suppressed by conventional metrics. In the cited literature, the term appears across image stitching quality assessment, light-field depth estimation, omnidirectional cross-reference evaluation, model stitching for representation comparison, evaluator-preference measurement in LLM agent systems, latent-space value estimation for diffusion alignment, and patch-based volume generation (Zhang et al., 2024, Zhou et al., 2022, Yu et al., 2019, Bansal et al., 2021, Liu, 1 Jul 2026, Go et al., 19 May 2026, Chamier et al., 18 Jun 2026). This suggests that the shared principle is methodological composition: partial evidence that would be inadequate in isolation is stitched into an evaluation object that is more sensitive to the structure actually relevant to the task.

1. General form and recurrent design pattern

Across the cited work, stitching serves one of a small number of technical roles. It may define the evaluation domain itself, as when corresponding EPIs are shifted and concatenated into a stitched EPI before candidate depths are scored; it may define the reference structure, as when orthogonal fisheye captures provide cross-reference observations of seam regions; it may define the evaluator, as when a diffusion backbone is attached to a reward-model tail; or it may define the experimental protocol, as in a four-phase isolation paradigm with fixed logging and versioning. A plausible implication is that a stitched evaluation protocol is best understood as a family of structured evaluation designs rather than a single canonical standard (Zhou et al., 2022, Yu et al., 2019, Go et al., 19 May 2026, Liu, 1 Jul 2026).

Setting	What is stitched	Primary evaluation target
Stitched-image IQA	Feature space or cross-reference observations	Correlation with subjective quality
Light-field depth estimation	Corresponding EPIs into SEPI or half-SEPI	Slope, depth, confidence
Neural representation comparison	Bottom of one model to top of another	Stitching penalty
LLM agent systems	Four experimental phases and manifests	Coupling metrics
Diffusion alignment	Diffusion backbone to reward-model tail	Value on noisy latents
Large-volume generation	Output patches or orthoslices	Perceptual quality and downstream IoU

Recurring requirements are explicit in several of these settings. For stitched-image quality assessment, desirable properties include high correlation with subjective scores, sensitivity to stitching-specific artifacts, robustness or stability, and a single usable indicator (Zhang et al., 2024). For LLM-agent evaluation, the emphasis shifts to reproducibility, comparability across evaluators and time points, and detection of measurement decay (Liu, 1 Jul 2026). For patch-based biomedical generation, the decisive requirement is that evaluation remain sensitive to downstream structural continuity rather than only global perceptual similarity (Chamier et al., 18 Jun 2026).

2. Full-reference stitched-image quality assessment

In image stitching research, the stitched evaluation protocol arises from the observation that standard FR and NR IQA metrics can contradict human judgments on stitched results because they mainly capture global distortion properties, whereas stitching artifacts are local, structured, and geometric: misalignment at overlaps, ghosting of moving objects or parallax, visible seams, local geometric distortions due to warps, and exposure or color inconsistencies between input views. SI-FID addresses this by retaining the standard Fréchet distance formula while replacing the original InceptionV3 feature extractor with a fine-tuned encoder trained by a SimSiam-style contrastive procedure on distorted images. The resulting score is selected as the altered FID whose feature space yields the highest PCC and SROCC against subjective stitched-image ratings; the best noise configuration is ColorJitter(brightness=0.5, hue=0.3). In the reported experiments, subjective evaluation used 14 computer vision researchers rating 160 stitched images on a continuous 0–100 scale under the same environment, and SI-FID achieved the highest PCC and SROCC on both test sets, with rank correlation at least 25% higher than other objective indicators (Zhang et al., 2024).

The formal distance remains

$$\operatorname{SI\!-\!FID}(\mathcal{X}_r,\mathcal{X}_s) = \|\mu_r-\mu_s\|_2^2 + \operatorname{Tr}\!\left(\Sigma_r+\Sigma_s-2(\Sigma_r\Sigma_s)^{1/2}\right),$$

but the feature mapping is changed from $f_{\text{orig}}$ to a fine-tuned $f_{\text{SI}}$. The protocol is therefore stitched at the feature-learning stage rather than at the metric formula itself. Its full-reference character is also explicit: the reference is the transformed image of the same scene, not the original source-image set or a seam mask.

A related but distinct full-reference design appears in omnidirectional cross-reference stitching. There, four dual-fisheye captures at $0^\circ$, $90^\circ$, $180^\circ$, and $270^\circ$ are acquired from the same camera position; when opposite views such as $0^\circ$ and $180^\circ$ are stitched, the orthogonal pair $90^\circ$ and $f_{\text{orig}}$0 acts as cross-reference ground truth for the stitching regions. The CROSS dataset contains 292 quaternions across 12 scene types, and the benchmark evaluates six stitching models with seven IQA metrics. The protocol is stitched because seam regions in the panorama are evaluated against reference observations that were not involved in stitching but cover the same physical scene with lower distortion in their central fisheye regions (Yu et al., 2019).

3. Geometric stitched domains and hypothesis evaluation

In light-field depth estimation, stitching defines the hypothesis-testing space rather than the final image. A 4D light field $f_{\text{orig}}$1 yields EPIs in which a 3D point traces a line whose slope is disparity-related. Conventional EPI-based estimation is limited by angular sparsity and discretization error. The stitched-EPI (SEPI) construction addresses this by identifying corresponding EPIs for the same 3D point across different angular rows, shifting them so that the candidate line aligns under a hypothesized slope $f_{\text{orig}}$2, and concatenating them into a stitched angular axis $f_{\text{orig}}$3. The shift and concatenation operators are

$f_{\text{orig}}$4

$f_{\text{orig}}$5

so a line that had only $f_{\text{orig}}$6 samples in a single EPI is evaluated with $f_{\text{orig}}$7 samples in SEPI. The candidate slope is selected by minimizing color variance along the stitched line,

$f_{\text{orig}}$8

with $f_{\text{orig}}$9 and confidence

$f_{\text{SI}}$0

The method further introduces half-SEPI, which restricts stitching to the non-occluded half of viewpoints for occlusion-edge pixels, and a propagation strategy for texture-less regions. On HCI Benchmark and HCI Blender, the reported average MSE improves over competing non-learning methods, and ablations show large degradation when SEPI, half-SEPI, or texture-less refinement is removed (Zhou et al., 2022).

This geometric interpretation matters because it clarifies that stitching is not merely output assembly. The protocol stitches evidence before evaluation. Candidate hypotheses are scored in a constructed domain whose sampling density is higher and whose visibility model can be selectively restricted. In that sense, SEPI is a stitched evaluation protocol in a literal geometric sense: the evaluand is a slope hypothesis, but the evidence for or against that hypothesis is assembled from multiple aligned EPIs.

4. Modular stitched protocols for representations and agent systems

Model stitching turns evaluation into a controlled interchangeability test between internal representations. Given frozen networks $f_{\text{SI}}$1 and $f_{\text{SI}}$2, a stitched model replaces the first $f_{\text{SI}}$3 layers of $f_{\text{SI}}$4 with $f_{\text{SI}}$5, inserts a low-capacity trainable stitcher $f_{\text{SI}}$6, and keeps $f_{\text{SI}}$7 fixed. The central quantity is the stitched loss

$f_{\text{SI}}$8

with $f_{\text{SI}}$9. For convolutional networks, the stitcher family is $0^\circ$0; for ViTs, it is a token-wise linear map. Only the stitcher is trained, typically with Adam and cosine learning-rate decay. The resulting stitching penalty measures whether two representations are compatible for the task and, because it is asymmetric, whether one representation can improve another. The paper uses this protocol to show low penalties between supervised and self-supervised models, negative penalties in “more is better” settings such as more data or more width, and a structural property termed stitching connectivity for independently trained SGD minima (Bansal et al., 2021).

EPC extends the stitched-protocol idea from networks to evaluator-driven agent systems. Its four-phase isolation paradigm defines pure Text, pure Visual, $0^\circ$1, and $0^\circ$2 phases, each running a TTRL update rule over a strategy distribution $0^\circ$3 with $0^\circ$4 strategies. The sampled strategy weight is updated by

$0^\circ$5

with $0^\circ$6 for evaluator wins and $0^\circ$7 for losses, followed by renormalization. Coupling is then measured by metrics such as

$0^\circ$8

along with JSD, and optionally ECE and Brier when ground-truth accuracy is available. EPC further requires a manifest JSON with protocol version, evaluator and executor metadata, configuration, task and strategy sets, and per-seed results; it introduces a versioning convention $0^\circ$9 and a time-bound Reference Snapshot v1.0 based on 122 experimental repetitions across multiple evaluator conditions. Here the stitched protocol is an RFC-style composition of modules, phases, update rules, metrics, and logging conventions (Liu, 1 Jul 2026).

A common misconception is that stitching in evaluation necessarily means image seam assessment. These two lines of work show otherwise. In one case, stitching connects model fragments to test representational equivalence; in the other, it standardizes an experimental pipeline so that third parties can reproduce and compare evaluator-preference measurements over time.

5. Stitched evaluators in generative modeling and large-volume synthesis

StitchVM applies stitching to the evaluator itself. A truncated diffusion backbone $90^\circ$0 is attached to a truncated reward-model tail $90^\circ$1 through a stitching layer $90^\circ$2, yielding

$90^\circ$3

where $90^\circ$4 is a noisy latent from the forward process $90^\circ$5, $90^\circ$6, $90^\circ$7. The objective is to approximate a value over noisy latents by regressing toward clean-image reward outputs using an L2 loss, while the paper argues that the learned standard value shares the leading-order gradient of the soft value used in reward-tilted alignment. The diffusion backbone is frozen; the stitch layer is initialized by a closed-form least-squares feature match and then lightly fine-tuned. The protocol is stitched because it constructs a latent-space evaluator by joining a noise-native backbone and a pretrained pixel-space reward tail, avoiding per-sample Tweedie-style or Monte Carlo approximation. Reported downstream effects include $90^\circ$8 faster DPS with roughly halved peak GPU memory and $90^\circ$9 faster DiffusionNFT (Go et al., 19 May 2026).

Patch-based volume generation exposes another sense in which stitched evaluation protocols are necessary. In large cryo-EM volumes, generators cannot process the full volume at once, so inference proceeds by cutting the volume into patches, generating patch outputs, and stitching those patches back together. The paper compares three assembly schemes: artifact-free tile-and-stitch with valid convolutions and heavy central cropping; padded convolutions with overlapping patches and uniform averaging; and valid convolutions with no cropping and no overlap. It evaluates both 2D and 3D cycleGANs, plus 2D orthoslice ensembling from three orthogonal directions. The central finding is that FID fails to detect subtle stitching artifacts that significantly impact downstream mitochondria segmentation: best FID checkpoints often do not correspond to best IoU, and only very poor FID values clearly correlate with degraded segmentation. The best reported IoU is 0.7383 for 3D tile-and-stitch, only marginally above 2D tile-and-stitch at 0.7374, leading to the conclusion that 3D can marginally outperform 2D on downstream tasks when stitching is artifact-free, but the gain barely justifies the added computational cost (Chamier et al., 18 Jun 2026).

These two cases are complementary. StitchVM stitches modules to produce a better evaluator in latent space; the cryo-EM study evaluates how the stitching of generated outputs affects downstream task performance and shows that perceptual metrics alone are insufficient. Together they illustrate that stitched evaluation can operate either before generation, at evaluator construction time, or after generation, at reconstruction-and-benchmark time.

6. Scope, limitations, and methodological significance

The literature also delineates the limits of stitched evaluation. SI-FID depends strongly on augmentation choice, is trained on DIR-D scenes at $180^\circ$0, and remains a full-reference metric requiring transformed images of the same scene (Zhang et al., 2024). Omnidirectional cross-reference stitching requires special four-orientation acquisition and sufficient calibration detail to map seam regions into orthogonal fisheye views (Yu et al., 2019). SEPI assumes Lambertian surfaces, regular camera geometry, and piecewise-linear depth variation in texture-less regions, and is computationally heavy on CPU (Zhou et al., 2022). EPC measurements are explicitly time-bound and subject to version decay as evaluator models and API gateways update (Liu, 1 Jul 2026). StitchVM inherits the biases of its base reward model and is sensitive to stitch location, with deeper reward-block insertion performing poorly (Go et al., 19 May 2026). Patch-based biomedical generation shows that even artifact-free assembly can be architecture-specific and that FID, especially with natural-image features, may not reflect the continuity properties that matter for scientific downstream tasks (Chamier et al., 18 Jun 2026).

A recurrent misconception is that a stitched protocol automatically yields a single universal scalar. Some protocols do aim at one primary indicator, as SI-FID explicitly does, but others are intentionally multi-output: EPC standardizes $180^\circ$1, JSD, ECE, Brier, zero-coupling rate, and bootstrap intervals; model stitching produces layerwise penalties rather than a single summary; cross-reference omnidirectional evaluation combines FR and NR metrics with subjective pairwise MOS. Another misconception is that stitching only improves evaluation by adding more data. In several of these works, the critical advance is not quantity but structure: occluded views are excluded in half-SEPI; only a low-capacity stitcher is trained in representation comparison; evaluator metadata and measurement dates are logged in EPC; only padding-free central crops are trusted in tile-and-stitch.

Taken together, the cited work supports a precise encyclopedic interpretation. A stitched evaluation protocol is a task-specific methodology in which evaluation reliability is improved by stitching together aligned partial evidence under explicit structural constraints. Depending on the domain, those constraints may be geometric, architectural, experimental, or statistical. The significance of the concept lies not in the word “stitched” itself, but in the consistent methodological claim behind these systems: when the failure mode is local, structured, modular, or time-varying, evaluation must be constructed to reflect that structure rather than averaged away by a global metric.