TuningIQA: Fine-Grained BIQA for Livestreaming

• TuningIQA is a fine-grained BIQA metric that predicts both overall quality scores and subtle pairwise preferences for optimized livestreaming camera tuning.
• It leverages the FGLive-10K dataset with multi-attribute and pairwise annotations to support precise camera parameter adjustments under real-world conditions.
• The approach combines human-aware feature extraction with graph-based parameter fusion, demonstrating improved performance over state-of-the-art IQA models.

Searching arXiv for the target paper and closely related BIQA / unified IQA-IAA work to ground the article. I’ll look up the TuningIQA paper and a few adjacent works on fine-grained BIQA, unified IQA/IAA, and explainable IQA. TuningIQA is a fine-grained blind image quality assessment (BIQA) metric for livestreaming camera tuning, introduced together with the FGLive-10K database to address a limitation of existing BIQA models: they typically predict only an overall coarse-grained quality score and therefore do not provide the fine-grained perceptual guidance required for precise camera parameter tuning in livestreaming workflows (Sheng et al., 25 Aug 2025). The framework is designed for scenarios in which camera parameters must be optimized automatically to improve user Quality of Experience (QoE), and it combines human-aware feature extraction with graph-based camera parameter fusion so that quality prediction can support both scalar score regression and fine-grained quality ranking under varying camera configurations (Sheng et al., 25 Aug 2025).

1. Problem setting and research context

TuningIQA is situated at the intersection of no-reference image quality assessment, fine-grained perceptual modeling, and camera control for livestreaming. Its central premise is that automatic livestreaming camera tuning requires more than a single overall quality estimate: the control loop must distinguish subtle perceptual differences that arise when exposure components, white balance, or image enhancement parameters are adjusted (Sheng et al., 25 Aug 2025).

The framework targets blind image quality assessment, meaning that quality must be inferred without access to a reference image. Within the paper’s problem formulation, this is not merely a regression task over overall Mean Opinion Score (MOS); it is also a ranking task over perceptually close alternatives generated by different camera parameter settings. The work therefore introduces both multi-attribute quality annotations and fine-grained pairwise preference annotations, explicitly tying quality estimation to tuning decisions rather than to passive benchmark prediction alone (Sheng et al., 25 Aug 2025).

This emphasis distinguishes TuningIQA from several adjacent lines of IQA research. UniQA studies unified vision-language pre-training for image quality and aesthetic assessment and adapts a pre-trained model using a lightweight adapter and multi-cue integration prompts (Zhou et al., 2024). Q-Adapt addresses explainable IQA with progressive instruction tuning for overall quality explanation and attribute-wise perception question answering (Lu et al., 2 Apr 2025). TATAR focuses on unified IQA and image aesthetic assessment through task-conditioned reasoning and asymmetric rewards (Yin et al., 20 Mar 2026). Refine-IQA develops a multi-stage reinforcement fine-tuning framework for perceptual IQA with explicit supervision of the “think” process (Jia et al., 4 Aug 2025). Test-time adaptation methods for blind IQA instead address distribution shift at inference time through auxiliary self-supervised objectives (Roy et al., 2023). These neighboring efforts show that recent IQA research has expanded beyond scalar prediction toward reasoning, perception enhancement, and adaptation; TuningIQA adds a distinct camera-tuning-oriented formulation grounded in fine-grained perceptual supervision (Sheng et al., 25 Aug 2025).

2. FGLive-10K: dataset design and annotation protocol

TuningIQA is built on FGLive-10K, described as a comprehensive fine-grained BIQA database containing 10,185 high-resolution images captured under varying camera parameter configurations across diverse livestreaming scenarios (Sheng et al., 25 Aug 2025). The images are all 1920×1080 and were filtered to ensure that at least one person appears. The scenarios comprise 555 distinct scenes covering e-commerce, entertainment, and creative content, all containing human subjects (Sheng et al., 25 Aug 2025).

Image acquisition used three custom livestreaming cameras, each calibrated to professional “reference” settings per scene by Field Application Engineers. Systematic deviations were then applied independently and jointly to seven settings: aperture, shutter speed, ISO, white balance, contrast, saturation, and sharpness (Sheng et al., 25 Aug 2025). A metadata subset, FGLive-p, retains the full 7-parameter EXIF-style metadata for 6,707 images, split into 5,559 train and 1,148 test examples, specifically to support parameter-aware modeling (Sheng et al., 25 Aug 2025).

The annotation design has two complementary components. First, FGLive-10K contains 50,925 multi-attribute quality annotations (Sheng et al., 25 Aug 2025). Twenty-five volunteers, including IQA researchers, photographers, and art students, scored each image under the ITU-R BT.500-15 protocol. Five attributes were rated on a 1–5 integer scale: overall quality, face quality, sharpness, exposure, and noise. Each image received at least 16 independent scores, and per-attribute MOS was computed as

$$s^{\text{attr}}=\frac{1}{N}\sum_{i=1}^N s_i^{\text{attr}} \quad (\text{attr}\in\{\text{overall,face,sharp,exp,noise}\}).$$

(Sheng et al., 25 Aug 2025)

Second, the dataset contains 19,234 fine-grained pairwise preference annotations, embedded in a total of 91,946 pairwise comparisons (Sheng et al., 25 Aug 2025). The initial preference label for a pair was defined from MOS as

$c^*_{pq}=\mathbb{I}\bigl(s_p> s_q\bigr).$

When the MOS difference satisfied $\Delta s=|s_p-s_q|\le 0.8$, the pair was re-rated $K$ times using three choices, $$\{0: I_p \text{ worse}, 0.5: \text{ equivalent}, 1: I_p \text{ better}\}$$, and the final preference became

$c_{pq}= \begin{cases} c^*_{pq}, & \Delta s>0.8,\[4pt] \dfrac{1}{K}\sum_{k=1}^K \psi_k(I_p,I_q), & \Delta s\le0.8. \end{cases}$

(Sheng et al., 25 Aug 2025)

The paper reports a scene-level split with 451 train scenes comprising 8,148 images and 72,843 pairs, and 104 test scenes comprising 2,037 images and 3,705 pairs (Sheng et al., 25 Aug 2025). It also reports that overall and face scores are highly correlated, with PLCC approximately $0.9$, while sharpness, exposure, and noise exhibit pairwise PLCC below $0.4$, indicating complementarity among non-overall perceptual dimensions (Sheng et al., 25 Aug 2025). Global annotator consistency is approximately PLCC $0.85$, but within the high-MOS range $4.2$–$c^*_{pq}=\mathbb{I}\bigl(s_p> s_q\bigr).$0 it drops to approximately $c^*_{pq}=\mathbb{I}\bigl(s_p> s_q\bigr).$1, which is presented as evidence for the need for pairwise refinement at fine granularity (Sheng et al., 25 Aug 2025).

Aspect	FGLive-10K specification
Images and scenes	10,185 images across 555 scenes
Annotation types	50,925 multi-attribute annotations and 19,234 fine-grained pairwise preference annotations
Metadata subset	FGLive-p: 6,707 images with full 7-parameter metadata

These design choices make FGLive-10K not only a BIQA dataset but also a structured substrate for quality-guided camera control. A plausible implication is that the dataset operationalizes a transition from passive image assessment to decision-oriented assessment, because the labels directly encode both absolute quality and actionable pairwise preferences under parameter perturbations (Sheng et al., 25 Aug 2025).

3. Human-aware feature extraction and graph-based parameter fusion

The TuningIQA model supports both single-image scoring and pairwise fine-grained comparison through two modules: Human-aware Feature Extraction (HFE) and, when metadata are available, Graph-based Camera Parameter Fusion (GCPF) (Sheng et al., 25 Aug 2025).

HFE begins by resizing the input image to $c^*_{pq}=\mathbb{I}\bigl(s_p> s_q\bigr).$2, applying a random crop to $c^*_{pq}=\mathbb{I}\bigl(s_p> s_q\bigr).$3, and using horizontal and vertical flips (Sheng et al., 25 Aug 2025). Human subjects are localized with a pre-trained Faster-R-CNN, which produces human bounding boxes $c^*_{pq}=\mathbb{I}\bigl(s_p> s_q\bigr).$4 (Sheng et al., 25 Aug 2025). The image is then processed by an EfficientNetV2-M backbone pre-trained on human aesthetics (TAD66K), producing a multi-scale feature map

$c^*_{pq}=\mathbb{I}\bigl(s_p> s_q\bigr).$5

(Sheng et al., 25 Aug 2025)

Human region features $c^*_{pq}=\mathbb{I}\bigl(s_p> s_q\bigr).$6 are extracted by ROI-Align over each $c^*_{pq}=\mathbb{I}\bigl(s_p> s_q\bigr).$7 with dimensionality $c^*_{pq}=\mathbb{I}\bigl(s_p> s_q\bigr).$8, then concatenated in broadcast form to every spatial location of $c^*_{pq}=\mathbb{I}\bigl(s_p> s_q\bigr).$9 to produce an enriched feature representation (Sheng et al., 25 Aug 2025). This enriched map is partitioned into nine overlapping regions around the human center, denoted $\Delta s=|s_p-s_q|\le 0.8$0. Each region is processed by a partition-wise residual block:

$\Delta s=|s_p-s_q|\le 0.8$1

where $\Delta s=|s_p-s_q|\le 0.8$2 is a $\Delta s=|s_p-s_q|\le 0.8$3 convolution with $\Delta s=|s_p-s_q|\le 0.8$4 output channels (Sheng et al., 25 Aug 2025).

The regional outputs are pooled to obtain a human summary feature $\Delta s=|s_p-s_q|\le 0.8$5, while the backbone also produces a global feature $\Delta s=|s_p-s_q|\le 0.8$6. These are fused through cross-attention:

$\Delta s=|s_p-s_q|\le 0.8$7

$\Delta s=|s_p-s_q|\le 0.8$8

The resulting $\Delta s=|s_p-s_q|\le 0.8$9 is described as quality-aware features fusing human-centric and global cues (Sheng et al., 25 Aug 2025). The emphasis on human regions is consistent with the dataset’s focus on livestreaming scenes that always contain people, and especially with the presence of a face-quality attribute (Sheng et al., 25 Aug 2025).

When camera metadata are available, GCPF augments these visual features with parameter-structured reasoning. The graph nodes are

$K$0

and the edge set $K$1 encodes four relations: visual-to-parameter cross-modal links, full connectivity of the exposure triangle ISO–shutter–aperture, full connectivity of the post-processing chain contrast–saturation–sharpness, and a white-balance-to-saturation edge representing color correlation (Sheng et al., 25 Aug 2025). Node encodings are

$K$2

(Sheng et al., 25 Aug 2025)

Graph reasoning stacks two graph attention layers:

$K$3

with four heads in the first layer and one head in the second. The fused visual embedding is the visual node output,

$K$4

(Sheng et al., 25 Aug 2025)

The overall architecture therefore has a dual inductive bias. HFE enforces human-aware visual prioritization, and GCPF enforces structured interactions among camera parameters. This suggests that TuningIQA is not merely a generic BIQA backbone retrained on livestreaming data; rather, it is engineered around the causal and perceptual structure of livestreaming capture pipelines (Sheng et al., 25 Aug 2025).

4. Prediction tasks and optimization objectives

TuningIQA uses two prediction heads, aligned with the two annotation regimes of FGLive-10K (Sheng et al., 25 Aug 2025). For multi-attribute regression, one MLP is learned for each attribute in $K$5:

$K$6

where $K$7 is either $K$8 or $K$9 depending on whether metadata are present (Sheng et al., 25 Aug 2025).

For fine-grained pairwise classification, two images $$\{0: I_p \text{ worse}, 0.5: \text{ equivalent}, 1: I_p \text{ better}\}$$0 are compared using a feature composition that includes both concatenation and a difference term:

$$\{0: I_p \text{ worse}, 0.5: \text{ equivalent}, 1: I_p \text{ better}\}$$1

(Sheng et al., 25 Aug 2025)

Training jointly optimizes a confidence-weighted regression loss and a fine-grained ranking loss. The regression loss is

$$\{0: I_p \text{ worse}, 0.5: \text{ equivalent}, 1: I_p \text{ better}\}$$2

where $$\{0: I_p \text{ worse}, 0.5: \text{ equivalent}, 1: I_p \text{ better}\}$$3 is the empirical variance of annotator scores on image $$\{0: I_p \text{ worse}, 0.5: \text{ equivalent}, 1: I_p \text{ better}\}$$4 (Sheng et al., 25 Aug 2025). The factor $$\{0: I_p \text{ worse}, 0.5: \text{ equivalent}, 1: I_p \text{ better}\}$$5 down-weights uncertain samples, directly linking supervision strength to annotation consistency.

The ranking loss is

$$\{0: I_p \text{ worse}, 0.5: \text{ equivalent}, 1: I_p \text{ better}\}$$6

(Sheng et al., 25 Aug 2025)

The total objective is a weighted sum,

$$\{0: I_p \text{ worse}, 0.5: \text{ equivalent}, 1: I_p \text{ better}\}$$7

(Sheng et al., 25 Aug 2025)

The implementation uses PyTorch on NVIDIA RTX4090, AdamW with cosine-annealing schedule, maximum learning rate $$\{0: I_p \text{ worse}, 0.5: \text{ equivalent}, 1: I_p \text{ better}\}$$8, batch size $$\{0: I_p \text{ worse}, 0.5: \text{ equivalent}, 1: I_p \text{ better}\}$$9 for both ranking pairs and single-image regression, and five epochs of end-to-end joint training (Sheng et al., 25 Aug 2025). The reported backbones are MobileNetV3-S with $c_{pq}= \begin{cases} c^*_{pq}, & \Delta s>0.8,\[4pt] \dfrac{1}{K}\sum_{k=1}^K \psi_k(I_p,I_q), & \Delta s\le0.8. \end{cases}$0M parameters and ResNet-50 with $c_{pq}= \begin{cases} c^*_{pq}, & \Delta s>0.8,\[4pt] \dfrac{1}{K}\sum_{k=1}^K \psi_k(I_p,I_q), & \Delta s\le0.8. \end{cases}$1M parameters (Sheng et al., 25 Aug 2025). Augmentation is restricted to random crop and flips, explicitly excluding quality-altering operations (Sheng et al., 25 Aug 2025).

The combined training formulation is significant because it couples absolute judgment and relative judgment in one metric. A plausible implication is that the model is trained to be simultaneously calibration-aware and decision-aware: regression aligns it with MOS, while pairwise supervision aligns it with the small comparative distinctions needed by a tuning controller (Sheng et al., 25 Aug 2025).

5. Empirical evaluation and ablation evidence

Evaluation follows a scene-level split with no overlap between train and test scenes, and uses three metrics: Spearman’s Rank Correlation Coefficient (SRCC), Pearson Linear Correlation Coefficient (PLCC), and fine-grained ranking accuracy (FG-ACC), defined as the fraction of pairs for which

$c_{pq}= \begin{cases} c^*_{pq}, & \Delta s>0.8,\[4pt] \dfrac{1}{K}\sum_{k=1}^K \psi_k(I_p,I_q), & \Delta s\le0.8. \end{cases}$2

(Sheng et al., 25 Aug 2025)

On the FGLive-10K test set, using a ResNet-50 backbone, the paper reports the following headline comparison:

Method	SRCC	PLCC	FG-ACC
MUSIQ	0.8662	0.8687	0.6861
LIQE	0.9235	0.9206	0.6533
Q-Align	0.8721	0.8315	0.6311
TuningIQA	0.9385	0.9364	0.7284

These figures are used to support the claim that TuningIQA significantly outperforms state-of-the-art BIQA methods in both score regression and fine-grained quality ranking (Sheng et al., 25 Aug 2025). The paper further states that, for each attribute, TuningIQA’s SRCC, PLCC, and FG-ACC outperform CLIP-IQA, MT-A, and SARQUE, with details in Table 3 of the paper (Sheng et al., 25 Aug 2025).

The ablations separate the contributions of HFE and GCPF. On FGLive-10K without metadata, a baseline without HFE attains SRCC $c_{pq}= \begin{cases} c^*_{pq}, & \Delta s>0.8,\[4pt] \dfrac{1}{K}\sum_{k=1}^K \psi_k(I_p,I_q), & \Delta s\le0.8. \end{cases}$3, PLCC $c_{pq}= \begin{cases} c^*_{pq}, & \Delta s>0.8,\[4pt] \dfrac{1}{K}\sum_{k=1}^K \psi_k(I_p,I_q), & \Delta s\le0.8. \end{cases}$4, and FG-ACC $c_{pq}= \begin{cases} c^*_{pq}, & \Delta s>0.8,\[4pt] \dfrac{1}{K}\sum_{k=1}^K \psi_k(I_p,I_q), & \Delta s\le0.8. \end{cases}$5, while adding HFE increases performance to SRCC $c_{pq}= \begin{cases} c^*_{pq}, & \Delta s>0.8,\[4pt] \dfrac{1}{K}\sum_{k=1}^K \psi_k(I_p,I_q), & \Delta s\le0.8. \end{cases}$6, PLCC $c_{pq}= \begin{cases} c^*_{pq}, & \Delta s>0.8,\[4pt] \dfrac{1}{K}\sum_{k=1}^K \psi_k(I_p,I_q), & \Delta s\le0.8. \end{cases}$7, and FG-ACC $c_{pq}= \begin{cases} c^*_{pq}, & \Delta s>0.8,\[4pt] \dfrac{1}{K}\sum_{k=1}^K \psi_k(I_p,I_q), & \Delta s\le0.8. \end{cases}$8 (Sheng et al., 25 Aug 2025). On FGLive-p with metadata, the baseline attains $c_{pq}= \begin{cases} c^*_{pq}, & \Delta s>0.8,\[4pt] \dfrac{1}{K}\sum_{k=1}^K \psi_k(I_p,I_q), & \Delta s\le0.8. \end{cases}$9, adding GCPF yields $0.9$0, and combining GCPF with HFE yields $0.9$1 (Sheng et al., 25 Aug 2025). The paper concludes that both HFE and GCPF bring consistent gains, especially for FG-ACC (Sheng et al., 25 Aug 2025).

For completeness, the paper also writes out the metric formulas. Spearman’s $0.9$2 is

$0.9$3

where $0.9$4 are rank differences, and PLCC is

$0.9$5

(Sheng et al., 25 Aug 2025)

The evaluation protocol emphasizes ranking as much as regression. This is notable because many IQA works report only correlation with MOS, whereas TuningIQA treats ranking accuracy as a first-class deployment metric. That design is consistent with its use in camera control, where the operative question is often which small adjustment is better rather than what absolute score each candidate deserves (Sheng et al., 25 Aug 2025).

6. Deployment in livestreaming camera tuning

The deployment formulation is explicit: at runtime, the system initializes camera parameters at scene start and iteratively captures a frame, extracts HFE features, optionally fuses current metadata through GCPF, predicts coarse attribute scores, evaluates candidate parameter adjustments through the fine-grained pairwise head, applies the preferred adjustment, and stops when no adjustment improves quality or after a fixed number of steps (Sheng et al., 25 Aug 2025).

The workflow is summarized in the paper as:

1. capture frame $0.9$6,
2. compute $0.9$7 by HFE,
3. if metadata are available, compute $0.9$8 through GCPF, otherwise set $0.9$9,
4. predict coarse scores $0.4$0,
5. generate small candidate adjustments $0.4$1,
6. estimate pairwise preference for each candidate,
7. select the $0.4$2 maximizing predicted preference,
8. update camera parameters,
9. terminate when there is no predicted improvement (Sheng et al., 25 Aug 2025).

The paper also provides practical performance notes. With the MobileNetV3-S backbone and GAT, inference latency is approximately $0.4$3–$0.4$4 ms per frame on RTX-class GPUs, corresponding to about $0.4$5–$0.4$6 FPS; on embedded GPUs such as Jetson Xavier, the expectation is $0.4$7–$0.4$8 FPS, which is described as still real-time for slow camera adjustments (Sheng et al., 25 Aug 2025). The model footprint is reported as $0.4$9M parameters for the MobileNet variant or $0.85$0M for the ResNet variant, plus GAT layers, fitting in $0.85$1–$0.85$2 GB GPU memory (Sheng et al., 25 Aug 2025). The paper further notes that new parameters such as HDR gain can be added as extra nodes in GCPF with minimal retraining (Sheng et al., 25 Aug 2025).

This operationalization clarifies that TuningIQA is not solely an offline benchmark model. It is a quality metric embedded in a control loop. In that respect, it differs from test-time adaptation work that updates a blind IQA model at inference time to handle distribution shift (Roy et al., 2023), and from explainable IQA systems that focus on text generation or attribute-wise question answering (Lu et al., 2 Apr 2025, Li et al., 4 Oct 2025). TuningIQA’s primary endpoint is camera parameter selection under livestreaming constraints (Sheng et al., 25 Aug 2025).

7. Position within the evolving IQA landscape

TuningIQA belongs to a broader shift in IQA research from monolithic scalar prediction toward richer task formulations. UniQA shows that quality and aesthetics can be jointly represented in a unified vision-language latent space and then adapted with a lightweight multi-cue integration adapter (Zhou et al., 2024). Q-Adapt organizes explainable IQA as a two-stage progressive instruction-tuning problem over quality explanation and attribute-wise perception question answering (Lu et al., 2 Apr 2025). TATAR argues that unified IQA and aesthetic assessment require task-conditioned reasoning styles and asymmetric rewards because IQA and IAA differ fundamentally in reasoning and optimization structure (Yin et al., 20 Mar 2026). Refine-IQA introduces multi-stage reinforcement fine-tuning with distortion perception tasks in Stage 1 and quality scoring with “think” supervision in Stage 2 (Jia et al., 4 Aug 2025). IQA-Select studies instruction-data redundancy and shows that clustering-based coreset selection can exceed full-data fine-tuning using only $0.85$3 selected data in Q-Bench and AesBench (Li et al., 4 Oct 2025).

Against that background, TuningIQA contributes a distinct problem definition: fine-grained BIQA for livestreaming camera tuning (Sheng et al., 25 Aug 2025). Its innovations are not framed around multimodal reasoning traces, instruction-following, or reinforcement fine-tuning, but around the structure of the livestreaming environment itself: human-centered visual content, explicit camera parameter metadata, pairwise perceptual refinement near high-quality regimes, and deployment within a tuning loop (Sheng et al., 25 Aug 2025).

A common misconception is that a strong overall BIQA regressor is sufficient for camera tuning. The TuningIQA formulation argues otherwise by constructing a dataset in which coarse overall scores are supplemented by attribute-specific MOS and fine-grained pairwise preferences, and by showing that dedicated modeling of human-aware cues and parameter relations improves fine-grained ranking performance (Sheng et al., 25 Aug 2025). Another possible misconception is that camera metadata are always necessary. The architecture is explicitly designed to operate with or without metadata: HFE alone supports image-only assessment, while GCPF is an optional extension for FGLive-p or other metadata-rich settings (Sheng et al., 25 Aug 2025).

In summary, TuningIQA defines a specialized but technically consequential extension of BIQA: a metric that predicts both coarse perceptual quality and subtle relative preference under controllable camera perturbations, grounded in a livestreaming-specific dataset and realized through human-aware feature extraction plus graph-based parameter fusion (Sheng et al., 25 Aug 2025). Its formulation suggests a broader direction for future IQA systems in which assessment models are trained not only to describe perception, but also to guide closed-loop visual optimization.