Quality Factor-Aware Time Predictor

• The topic defines quality factor-aware time predictors as models that use quality proxies (e.g., Octree depth, JPEG QF) to inform time-related predictions.
• It covers dual approaches including online control for delay stability and structured regression that integrates quality signals into operator-level performance.
• Empirical studies demonstrate that incorporating quality factors significantly improves prediction accuracy and system performance across AR, DL training, manufacturing, and software execution.

The expression quality factor-aware time predictor can be understood as an Editor’s term for a class of models and controllers that make time-related decisions or predictions using an explicit quality variable, quality proxy, or quality-conditioned representation. In the recent literature, this role is instantiated in several distinct ways: Octree depth as a quality-control knob for bounded-delay augmented-reality visualization, JPEG Quality Factor as a learned reference-free artifact proxy that can condition runtime regressors, floating-point precision as a determinant of distributed training iteration time, phase-aligned frequency structure as a predictor of future manufacturing quality trajectories, and LLVM-derived execution-quality indicators as predictors of software execution time (Lee et al., 2022, Cui et al., 2024, Kang et al., 17 Apr 2026, Luo et al., 30 Jul 2025, Xu et al., 17 Mar 2025). The unifying theme is that time is not treated as a function of input size alone; it is modeled jointly with fidelity, degradation severity, numerical precision, or process-specific quality dynamics.

1. Conceptual scope and representative forms

The literature suggests two adjacent senses of the concept. In one sense, the predictor estimates or controls execution time, training time, or delay from a quality-related factor. In the other, it predicts future quality over time by explicitly modeling quality-generating temporal factors such as lag, periodicity, and shared frequency structure. Both senses are present in the cited work.

Domain	Quality factor or proxy	Time-related target
AR visualization	Octree depth $d(t)$	queue backlog / delay stability
Reference-free IQA	JPEG Quality Factor or predicted QF map	runtime-conditioning signal
Distributed DL training	FP32, FP16, mixed precision	single-iteration training time
Multi-process manufacturing	phase alignment and frequency components	future quality over horizon $H$
LLVM timing analysis	cache, branch, and control-flow quality indicators	average core execution time

In "Quality-Aware Real-Time Augmented Reality Visualization under Delay Constraints" (Lee et al., 2022), quality is the visualization quality induced by Octree depth. In "Reference-Free Image Quality Metric for Degradation and Reconstruction Artifacts" (Cui et al., 2024), JPEG Quality Factor is treated as a scalar proxy for artifact severity and is explicitly proposed as a conditioning variable for a downstream time regressor. In "Training Time Prediction for Mixed Precision-based Distributed Training" (Kang et al., 17 Apr 2026), floating-point precision acts as the quality factor because it changes compute throughput, communication volume, and the effective performance envelope. In "PAF-Net: Phase-Aligned Frequency Decoupling Network for Multi-Process Manufacturing Quality Prediction" (Luo et al., 30 Jul 2025), quality prediction is itself a time-series forecasting problem whose structure is governed by lagged and frequency-specific quality factors. In "PrETi: Predicting Execution Time in Early Stage with LLVM and Machine Learning" (Xu et al., 17 Mar 2025), timing is conditioned on features that quantify the quality of locality, predictability, and control flow.

2. Mathematical formulations of quality–time coupling

A canonical formulation appears in the AR scheduling model, which poses time-average quality maximization under a delay constraint: $$\max \quad \lim_{t\rightarrow\infty}\frac{1}{t}\sum_{\tau=0}^{t-1} P_{a}(d(\tau))$$ subject to

$$\lim_{t\rightarrow\infty}\frac{1}{t}\sum_{\tau=0}^{t-1} Q(\tau)< \infty.$$

Here $P_a(d(\tau))$ is visualization quality, and $Q(\tau)$ is a queue backlog that serves as a proxy for delay. The per-slot Lyapunov control law is

$$d^{*}(t)\leftarrow \arg\max_{d(t)\in\mathcal{R}} \Big[ V\cdot p_{a}(d(t)) - Q(t)a(d(t)) \Big],$$

with $V$ as the quality–delay tradeoff coefficient (Lee et al., 2022). This formulation makes the quality factor operational: a deeper Octree increases quality and simultaneously increases workload $a(d(t))$, so the controller selects quality as a congestion-weighted scheduling decision.

A different but closely related structure appears in precision-aware training-time prediction. There, the paper decomposes single-iteration time into computation and communication components: $$T(d, t, p) = T_{\text{comp}}(d, t, p) + T_{dp}(d) + T_{tp}(t) + T_{pp}(p).$$ Precision is part of the job configuration and is translated into an operator-level precision map, which determines both profiled operator latency and communicated tensor size. Communication overheads follow the standard volume-over-bandwidth form,

$H$0

while pipeline overhead is modeled as

$H$1

This suggests a general abstraction in which the time predictor evaluates a transformed execution graph under a quality factor and then aggregates compute, communication, and overhead contributions (Kang et al., 17 Apr 2026).

Taken together, these formulations show two principal mathematical regimes. One is online control, in which a quality factor is chosen subject to delay stability. The other is structured regression, in which a quality factor modifies operator latency, data size, or cross-process dependence, and time is predicted from the resulting structure.

3. Quality-factor representations

The most explicit learned quality representation is the QF Predictor, a lightweight, fully convolutional network with seven convolutional layers, two early max-pooling layers, Batch Normalization and ReLU after each convolution, two final $H$2 convolutions, and a Sigmoid output constrained to $H$3 (Cui et al., 2024). Rather than outputting a single scalar, it produces a QF map: $H$4 where $H$5 is a JPEG-compressed image and $H$6 is the normalized label. A pooled scalar score can then be written as

$H$7

The network is trained in a self-supervised manner by predicting the JPEG Quality Factor of randomly compressed patches. The final model uses regression with MSE loss,

$H$8

because regression captures subtle differences such as QF 90 versus 95 more effectively than the exploratory classification setup. The training distribution is intentionally skewed toward the high-QF regime via a logarithmic weighting, so that the model learns subtle perceptual degradations instead of relying mainly on obvious blocking artifacts (Cui et al., 2024).

This representation matters for time prediction because the paper explicitly describes several integration strategies. The predicted scalar $H$9, summary statistics of the QF map, or penultimate-layer embeddings can be fed into a second model

$$\max \quad \lim_{t\rightarrow\infty}\frac{1}{t}\sum_{\tau=0}^{t-1} P_{a}(d(\tau))$$0

where $$\max \quad \lim_{t\rightarrow\infty}\frac{1}{t}\sum_{\tau=0}^{t-1} P_{a}(d(\tau))$$1 denotes algorithm hyperparameters. The same paper also emphasizes that the learned quality notion generalizes beyond JPEG: predicted QF decreases monotonically with increasing blur or noise, and a grayscale version can proxy MRI undersampling severity, especially when computed on central patches rather than whole-image averages (Cui et al., 2024). A plausible implication is that a quality factor-aware time predictor need not use a handcrafted quality variable; it may instead use a learned, spatially resolved quality field.

4. Domain-specific predictor architectures

In distributed training, the decisive architectural idea is operator-level precision awareness. The model graph is extracted with torch.fx, partitioned into GPU-specific subgraphs under data, tensor, and pipeline parallelism, and then annotated with an operator-level precision map. For mixed precision, this map is obtained by hooking into torch.amp and recording the actual casted precision of each operator. Forward and backward operator latencies are then profiled at the actual tensor shapes and precisions used, and the predictor sums these measurements across subgraphs while also recomputing communication volumes under the same precision assumptions (Kang et al., 17 Apr 2026). The resulting design is neither purely analytical nor purely learned; it is a hybrid analytical plus profiling-based predictor.

PrETi follows a different architecture. It compiles C code to LLVM IR, instruments the IR into exec_flow.ll, branch.ll, and dcache.ll, executes the instrumented IR with lli, and then extracts feature vectors containing dynamic counts of executed LLVM IR instructions, load_hit, load_miss, store_hit, store_miss, br_hit, br_miss, bb_jump, inst_miss, and data-volume features for memcpy, memset, calloc, and malloc. These features are used by Linear Regression, Huber Regression, Random Forest, and a shallow MLP to predict average core execution time on an Infineon Aurix TC387 under cold-cache conditions (Xu et al., 17 Mar 2025). In this setting, the quality factors are not perceptual or numerical; they are locality quality, branch predictability, and control-flow regularity.

PAF-Net addresses a third regime: future quality forecasting in multi-process manufacturing. Its architecture begins with phase-correlation alignment,

$$\max \quad \lim_{t\rightarrow\infty}\frac{1}{t}\sum_{\tau=0}^{t-1} P_{a}(d(\tau))$$2

followed by neighbor selection, lag estimation, and circular alignment of neighboring processes. The aligned series are decomposed with DCT, partitioned into temporal patches of length $$\max \quad \lim_{t\rightarrow\infty}\frac{1}{t}\sum_{\tau=0}^{t-1} P_{a}(d(\tau))$$3, processed by frequency-independent patch attention, and then fused by frequency-decoupled cross attention before an MLP predicts the next $$\max \quad \lim_{t\rightarrow\infty}\frac{1}{t}\sum_{\tau=0}^{t-1} P_{a}(d(\tau))$$4 steps (Luo et al., 30 Jul 2025). Here the quality factors are phase, periodicity, and shared-band dependence; time awareness enters through explicit lag modeling and multi-step forecasting.

5. Empirical behavior and trade-offs

The empirical record shows that ignoring the quality factor is often catastrophic. In AR scheduling, the max-depth baseline causes the queue backlog to diverge and leads to queue overflow, whereas min-depth drives the queue backlog to 0 but sacrifices quality. The Lyapunov-based policy maintains a bounded queue, sustains high-quality depth up to an “optimized point” around 400 time units, and then reduces depth adaptively when backlog grows (Lee et al., 2022). This directly falsifies the common assumption that maximal fidelity is always compatible with real-time constraints.

In distributed training, the effect of precision is quantified sharply. On 8 NVIDIA H100 GPUs, training LLaMA 3.1–8B under FP32, FP16, and mixed precision produces training-time variations of about $$\max \quad \lim_{t\rightarrow\infty}\frac{1}{t}\sum_{\tau=0}^{t-1} P_{a}(d(\tau))$$5 over the minimum. When FP32-based predictors are applied without precision awareness, errors grow dramatically: NeuSight incurs an increase of approximately 130.55% in MAPE under mixed or unseen precisions, and vTrain reaches 147.85% MAPE. The proposed precision-aware predictor reduces average MAPE to about 9.8% for mixed precision and about 10.64% for FP16, corresponding to an approximately $$\max \quad \lim_{t\rightarrow\infty}\frac{1}{t}\sum_{\tau=0}^{t-1} P_{a}(d(\tau))$$6 improvement in prediction accuracy over the baselines across precision and parallelism configurations (Kang et al., 17 Apr 2026).

The QF Predictor shows that a learned quality map can be stable enough to serve as a runtime-conditioning signal. It ranks dataset-level quality in the order LIVE $$\max \quad \lim_{t\rightarrow\infty}\frac{1}{t}\sum_{\tau=0}^{t-1} P_{a}(d(\tau))$$7 Flickr1024 $$\max \quad \lim_{t\rightarrow\infty}\frac{1}{t}\sum_{\tau=0}^{t-1} P_{a}(d(\tau))$$8 ImageNet, with predicted mean QF values $$\max \quad \lim_{t\rightarrow\infty}\frac{1}{t}\sum_{\tau=0}^{t-1} P_{a}(d(\tau))$$9, $$\lim_{t\rightarrow\infty}\frac{1}{t}\sum_{\tau=0}^{t-1} Q(\tau)< \infty.$$0, and $$\lim_{t\rightarrow\infty}\frac{1}{t}\sum_{\tau=0}^{t-1} Q(\tau)< \infty.$$1, respectively. It also exhibits monotonic QF decrease as blur or noise intensity increases, and training converges roughly after about 500 epochs (Cui et al., 2024).

PAF-Net reports systematic forecasting gains on four real-world datasets. Against ten baselines, it achieves 7.06% lower MSE and 3.88% lower MAE on average, with average rank 1.00 in both MSE and MAE across four datasets and four horizons. Ablation results identify the largest degradation when frequency-independent decomposition is removed, while removing phase alignment or replacing frequency-decoupled cross attention also causes consistent deterioration (Luo et al., 30 Jul 2025). This indicates that, in manufacturing quality forecasting, the decisive time predictors are not merely temporal encoders but encoders that preserve phase and frequency structure.

PrETi demonstrates that quality-aware execution features can substitute for costly hardware tracing. Its Random Forest achieves 11.98% average APE and 12.11% average sAPE on the CATREEN test set, outperforming CATREEN at 17.41% average APE and ITHEMAL at 21.02% average APE, while reducing timing-analysis latency from 2–5 minutes per module to about 5 seconds, approximately a $$\lim_{t\rightarrow\infty}\frac{1}{t}\sum_{\tau=0}^{t-1} Q(\tau)< \infty.$$2 speedup (Xu et al., 17 Mar 2025).

6. Limitations, misconceptions, and extensions

Several misconceptions are directly contradicted by the literature. One is that a quality proxy can safely be optimized in isolation. The QF Predictor paper shows that when QF is used as a loss term for reconstruction, optimizing for QF alone can introduce unnatural colors or artifacts in RGB and can conflict with data consistency in MRI (Cui et al., 2024). Another is that a single static computation graph is sufficient for training-time prediction; the precision-aware study shows that mixed precision changes kernel latencies, communication volumes, and tensor-core utilization enough to make FP32-only models unreliable (Kang et al., 17 Apr 2026).

The limits of calibration are also domain-specific. The MRI case shows that whole-image QF averages can be biased by large dark background regions, so region-focused QF is preferred. The precision-aware training predictor is evaluated on a single node with eight homogeneous H100 GPUs interconnected by NVLink and focuses on per-iteration time rather than full time-to-convergence. PrETi is tied to a specific hardware/compiler configuration, models aCET rather than wCET, and does not support many C standard library functions such as fopen (Cui et al., 2024, Kang et al., 17 Apr 2026, Xu et al., 17 Mar 2025).

A broader synthesis nonetheless emerges. The cited work suggests a reusable pattern: define a quality factor $$\lim_{t\rightarrow\infty}\frac{1}{t}\sum_{\tau=0}^{t-1} Q(\tau)< \infty.$$3, derive how $$\lim_{t\rightarrow\infty}\frac{1}{t}\sum_{\tau=0}^{t-1} Q(\tau)< \infty.$$4 transforms workload, latency, or inter-process dependence, and then predict time from the transformed system. Depending on the domain, $$\lim_{t\rightarrow\infty}\frac{1}{t}\sum_{\tau=0}^{t-1} Q(\tau)< \infty.$$5 may be a scalar control knob such as Octree depth or precision, a learned perceptual variable such as predicted JPEG QF, or a structured latent object such as a phase-aligned frequency decomposition. The resulting predictor may be an online scheduler, a hybrid analytical-profiler, a neural forecast model, or a classical regressor. What makes it quality factor-aware is not the model family itself, but the explicit treatment of quality as a first-class determinant of temporal behavior (Lee et al., 2022, Luo et al., 30 Jul 2025).