Attention-Guided Seed Scorer

• Attention-guided seed scorer is a methodological pattern that uses attention-derived signals to prioritize candidate seeds before more computationally expensive downstream processing.
• In applications like 3D point-cloud detection, text-to-image diffusion, and medical imaging, these mechanisms leverage strategies such as dual query selection, core-token attention, and iterative seed harvesting to enhance accuracy.
• The approach consistently employs a coarse-to-fine filtering process, using early attention cues to reduce candidate complexity and improve overall performance.

Attention-guided seed scorer denotes a family of mechanisms that rank, select, or refine candidate “seeds” by using attention-derived evidence. In the cited literature, the term spans several technically distinct objects: object queries or “seeds” in 3D point-cloud detection, random seeds in text-to-image diffusion, and high-confidence seed samples in weakly supervised chest-radiograph analysis. A broader reading of adjacent work suggests that attention-guided scoring can also be interpreted through class-specific self-attention aggregation in Vision Transformers and uncertainty-guided spatial emphasis in segmentation, although those two formulations are not introduced primarily as seed scorers (Liu et al., 2024, Zhang et al., 19 May 2026, Tang et al., 2018, Leem et al., 2024, Ehab et al., 12 Apr 2026).

1. Scope of the term

A common source of confusion is that the word “seed” does not denote a single mathematical object across these papers. In SEED for 3D DETR, queries are also referred to as “seeds,” and the central problem is how to obtain appropriate object queries under high sparsity and uneven distribution of point clouds (Liu et al., 2024). In Attention-Based Seed Selection (ABSS), the seed is the random initialization used by a text-to-image diffusion model, and the core problem is that different initial seeds often yield large variations in image quality and prompt-image alignment (Zhang et al., 19 May 2026). In Attention-Guided Curriculum Learning (AGCL), seeds are high-confidence images selected for joint multi-label classification and weakly supervised localization of thoracic diseases, where image-level disease labels and a subset of disease severity-level information are available (Tang et al., 2018).

Setting	Seed definition	Attention-guided signal
3D point-cloud DETR	Queries / seeds	Confidence score, quality score, deformable grid attention
Text-to-image diffusion	Random seeds	Cross-attention or self-attention to core tokens
Weakly supervised chest radiographs	Seed samples	Classification probabilities and class-conditional heatmaps
ViT explainability	Not a seed selector in the narrow sense	Gradients guided by normalized self-attention scores
Plant seedling segmentation	Not a seed selector in the narrow sense	Uncertainty-guided dual attention and entropy weighting

This diversity means that “attention-guided seed scorer” is best understood as a methodological pattern rather than a single algorithm. The pattern is the use of an attention-derived or attention-modulated signal to prioritize candidates before more expensive or more consequential downstream computation.

2. Query scoring in 3D point clouds: SEED

SEED, “A Simple and Effective 3D DETR in Point Clouds,” identifies two main challenges for DETR-based 3D detection: obtaining appropriate object queries and implementing effective query interaction by exploiting the rich geometric structure of point clouds. Its solution combines a dual query selection (DQS) module with a deformable grid attention (DGA) module (Liu et al., 2024).

DQS is explicitly coarse-to-fine. In the foreground query selection stage, flattened BEV features $$\bm{F}_{bev} \in \mathbb{R}^{(H \times W) \times C}$$ are paired with predicted confidence scores $\bm{S}_{bev}$, and the model selects top candidates

$$\bm{Q}_c = \mathrm{Top}_{N_c}(\bm{F}_{bev}, \bm{S}_{bev}), \qquad N_c = H \times W \times r.$$

These coarse queries are then enhanced by one layer of the transformer decoder,

$$\bm{Q}_c' = \mathrm{Decoder}(\bm{Q}_c, \bm{F}_{bev}),$$

so that the high-recall pool already benefits from query interaction.

The fine stage predicts, for each coarse query, a classification score $\bm{S}_c$, a localization score $\bm{S}_l$ interpreted as predicted 3D IoU with ground truth, and a regression output $\bm{B}_c$. Query quality is then fused as

$$\bm{S}_q^i = \begin{cases} (\bm{S}_c^{i})^{1-\beta} \cdot (\bm{S}_l^{i})^{\beta}, & \text{if } \bm{S}_c^{i} > \tau \ \bm{S}_c^{i}, & \text{otherwise.} \end{cases}$$

The final proposals are the top $N_f$ boxes under $\bm{S}_q$,

$\bm{S}_{bev}$0

followed by

$\bm{S}_{bev}$1

In this formulation, confidence suppresses background queries, while localization quality suppresses poorly localized but high-objectness proposals. The result is an explicitly quality-aware seed scorer.

DGA addresses the second difficulty, namely how the selected seeds interact with point-cloud geometry. Each reference box is uniformly divided into a $\bm{S}_{bev}$2 grid, and the model predicts offsets for flexible sampling around those reference points. The enhanced query is computed as

$\bm{S}_{bev}$3

where $\bm{S}_{bev}$4. The grid contributes regular geometric cues of the object box, while offsets provide deformability. The training objective further introduces Quality-Aware Hungarian Matching, where the classification cost uses quality scores $\bm{S}_{bev}$5 rather than objectness alone, and the total Hungarian cost includes this term together with standard regression and GIoU terms.

The reported ablations make the seed-scoring role concrete. DQS outperforms one-step or heatmap-based query selection such as TransFusion and ConQueR by 1.4–1.6 mAPH/L2 on Waymo. DGA outperforms vanilla deformable attention and box-attention by +0.7 mAPH/L2. End-to-end, SEED-L reaches 75.5 / 73.5 mAP/mAPH (L2) on Waymo val with 100% train, and SEED reaches 71.2 NDS and 66.6 mAP on nuScenes.

3. Random-seed ranking in text-to-image diffusion: ABSS

ABSS, “Boosting Text-to-Image Diffusion Models via Core Token Attention-Based Seed Selection,” studies the seed effect in text-to-image generation and argues that early attention dynamics over prompt core tokens strongly predict final generation quality. Yunzhe Zhang, Hongfu Liu, and Pengyu Hong of Brandeis University introduce a training-free, plug-and-play, inference-time method that scores candidate seeds without finetuning and without altering the initial noise (Zhang et al., 19 May 2026).

The method begins with core-token identification. Core tokens are defined as the principal subject nouns, the content-bearing words expected to be dominant in the image. For a pool of $\bm{S}_{bev}$6 random seeds, the model runs only the first $\bm{S}_{bev}$7 denoising steps and extracts attention maps. In the U-Net setting, a denoising-step attention map is written as

$\bm{S}_{bev}$8

After aggregation across blocks and heads and temperature sharpening, the core-token concentration score is

$\bm{S}_{bev}$9

For DiT backbones, the paper instead averages self-attention weights from all image tokens to each core text token, smooths over text with a 1D Gaussian, and then averages across core tokens. In both cases, seeds are ranked by $$\bm{Q}_c = \mathrm{Top}_{N_c}(\bm{F}_{bev}, \bm{S}_{bev}), \qquad N_c = H \times W \times r.$$0, the top-$$\bm{Q}_c = \mathrm{Top}_{N_c}(\bm{F}_{bev}, \bm{S}_{bev}), \qquad N_c = H \times W \times r.$$1 are retained, and the remainder are discarded. The method does not rely on a fixed accept/reject threshold.

ABSS therefore operationalizes an attention-guided seed scorer in the most literal ranking sense: the seed is not a proposal or training sample but a stochastic generative initialization. Only the best candidates are fully denoised. The paper evaluates HPS, ImageReward, PickScore, CLIP Score, and NFE. On the SD 1.4/InitNO setting reported in Table 1, Random yields HPS 0.2711, IR 0.0612, PickScore 21.7945, CLIP 0.2721, and NFE 50, while ABSS yields HPS 0.2782, IR 0.1972, PickScore 21.8881, CLIP 0.2761, and NFE 73. The reported robustness result is NDCG $$\bm{Q}_c = \mathrm{Top}_{N_c}(\bm{F}_{bev}, \bm{S}_{bev}), \qquad N_c = H \times W \times r.$$2 for agreement with ground-truth human preference seed ranking, and the method remains effective even with up to 50% random noise in core-token identification.

A common misconception is that ABSS is an attention-guidance method in the sense of modifying diffusion trajectories. It is not described that way. The method is training-free, requires no model alteration, does not alter the initial noise, and functions as a lightweight pre-selection add-on.

4. Seed samples in weakly supervised thoracic disease analysis: AGCL

AGCL, “Attention-Guided Curriculum Learning for Weakly Supervised Classification and Localization of Thoracic Diseases on Chest Radiographs,” uses the term “seed” for high-confidence training images that most unambiguously display disease characteristics in both classification and localization. The framework combines curriculum learning, attention-guided seed selection, and iterative refinement under weak supervision, using only image-level disease labels and disease severity-level information for a subset (Tang et al., 2018).

Seed selection occurs in two ways. First, all images labeled with severe or moderate severity for a given disease are included as initial seeds, denoted $$\bm{Q}_c = \mathrm{Top}_{N_c}(\bm{F}_{bev}, \bm{S}_{bev}), \qquad N_c = H \times W \times r.$$3. Second, images without explicit disease severity labels can also become seeds if the current CNN classifies them with high confidence, measured by output probability greater than threshold $$\bm{Q}_c = \mathrm{Top}_{N_c}(\bm{F}_{bev}, \bm{S}_{bev}), \qquad N_c = H \times W \times r.$$4, with the example $$\bm{Q}_c = \mathrm{Top}_{N_c}(\bm{F}_{bev}, \bm{S}_{bev}), \qquad N_c = H \times W \times r.$$5, and if they are correctly classified as positive; these form $$\bm{Q}_c = \mathrm{Top}_{N_c}(\bm{F}_{bev}, \bm{S}_{bev}), \qquad N_c = H \times W \times r.$$6. The seed set is

$$\bm{Q}_c = \mathrm{Top}_{N_c}(\bm{F}_{bev}, \bm{S}_{bev}), \qquad N_c = H \times W \times r.$$7

Attention enters through class activation maps. For each selected seed image, the CNN produces a class-specific heatmap

$$\bm{Q}_c = \mathrm{Top}_{N_c}(\bm{F}_{bev}, \bm{S}_{bev}), \qquad N_c = H \times W \times r.$$8

where $$\bm{Q}_c = \mathrm{Top}_{N_c}(\bm{F}_{bev}, \bm{S}_{bev}), \qquad N_c = H \times W \times r.$$9 is the class weight in the last fully connected layer and $$\bm{Q}_c' = \mathrm{Decoder}(\bm{Q}_c, \bm{F}_{bev}),$$0 is the activation at location $$\bm{Q}_c' = \mathrm{Decoder}(\bm{Q}_c, \bm{F}_{bev}),$$1 in feature map $$\bm{Q}_c' = \mathrm{Decoder}(\bm{Q}_c, \bm{F}_{bev}),$$2. These seed heatmaps then supervise a second path in a two-path multi-task architecture. The classification path performs the disease-prediction task, while the attention path regresses toward the previously computed seed heatmaps using a channel-wise smooth L1 loss:

$$\bm{Q}_c' = \mathrm{Decoder}(\bm{Q}_c, \bm{F}_{bev}),$$3

The final loss is

$$\bm{Q}_c' = \mathrm{Decoder}(\bm{Q}_c, \bm{F}_{bev}),$$4

with $$\bm{Q}_c' = \mathrm{Decoder}(\bm{Q}_c, \bm{F}_{bev}),$$5 set empirically to 0.005.

The seed scorer here is therefore neither query ranking nor random-seed ranking. It is an iterative mechanism that identifies highly confident seed images and then uses their class-conditional heatmaps to guide the next round of feature learning. After training on the initial seed set, the model is re-applied to all positive-labeled images to harvest new high-confidence seeds, refine heatmaps, and repeat the process until no additional seeds are found or validation improvements plateau.

The reported gains are substantial. Baseline average AUC is 0.7708; AGCL after two iterations reaches average AUC 0.8027, which the summary describes as approximately 3.2 percentage points over baseline and at least 5.7% over the older state of the art. For localization on the test set at IoBB threshold 0.25, AGCL reports Recall 0.73 and Precision 0.44, with the paper summary also stating 7%/11% increases in Recall/Precision compared to the state of the art.

5. Adjacent scoring formulations: self-attention explanation and uncertainty-guided emphasis

Although not introduced primarily as seed scorers, two adjacent papers make explicit how attention can be converted into a scoring or weighting mechanism for spatial evidence. “Attention Guided CAM: Visual Explanations of Vision Transformer Guided by Self-Attention” develops a ViT visualization method that selectively aggregates gradients directly propagated from the classification output to each self-attention and guides those gradients with normalized self-attention scores. The feature map is defined as $$\bm{Q}_c' = \mathrm{Decoder}(\bm{Q}_c, \bm{F}_{bev}),$$6 with sigmoid normalization, and the final map is

$$\bm{Q}_c' = \mathrm{Decoder}(\bm{Q}_c, \bm{F}_{bev}),$$7

The paper stresses that sigmoid normalization does not affect actual ViT operation because it is applied only after inference for visualization. Quantitatively, it reports pixel accuracy 0.7341 on ILSVRC, IoU 0.5212, F1/Dice 0.6515, and ABPC 0.3691, outperforming Attention Rollout and LRP-based alternatives in the stated comparisons (Leem et al., 2024).

UGDA-Net, “Uncertainty-Guided Attention and Entropy-Weighted Loss for Precise Plant Seedling Segmentation,” uses uncertainty-guided dual attention rather than discrete seed selection. Given a feature map tensor $$\bm{Q}_c' = \mathrm{Decoder}(\bm{Q}_c, \bm{F}_{bev}),$$8, it computes channel attention $$\bm{Q}_c' = \mathrm{Decoder}(\bm{Q}_c, \bm{F}_{bev}),$$9, spatial attention $\bm{S}_c$0, and an uncertainty map from channel-wise standard deviation,

$\bm{S}_c$1

These are combined as

$\bm{S}_c$2

and the output is

$\bm{S}_c$3

with learnable scalar $\bm{S}_c$4 initialized as 0.1. The loss further emphasizes uncertain pixels through Shannon entropy,

$\bm{S}_c$5

with empirical $\bm{S}_c$6. The paper reports Dice improvements from 0.4233 to 0.5159 for U-Net and from 0.3519 to 0.4840 for LinkNet, described also as 9.3% and 13.2% above baseline, together with reduced false positives at the seedling-soil boundary (Ehab et al., 12 Apr 2026).

A plausible implication is that the seed-scoring idea extends beyond explicit top-$\bm{S}_c$7 selection. In these adjacent works, attention and uncertainty act as continuous scoring fields over patches or pixels, thereby prioritizing regions rather than discrete candidates.

6. Recurrent design patterns and conceptual distinctions

Taken together, these papers suggest a recurring design pattern: attention-guided seed scoring is usually a pre-allocation mechanism for limited downstream computation. In SEED, high-recall candidates are narrowed to high-quality object queries before later decoding and geometric aggregation (Liu et al., 2024). In ABSS, only the top-ranked random seeds are denoised to completion (Zhang et al., 19 May 2026). In AGCL, only highly confident images and their heatmaps become privileged supervision for subsequent training iterations (Tang et al., 2018).

A second recurrent pattern is coarse-to-fine screening. SEED first uses predicted confidence scores and then estimated quality scores. ABSS first performs partial denoising and only then commits to full denoising for the top-$\bm{S}_c$8 seeds. AGCL begins with severe and moderate disease samples and then expands the seed pool through iterative harvesting of high-confidence positives. This suggests that attention-guided seed scoring is especially useful when exhaustive evaluation is computationally prohibitive or weak supervision makes direct localization unreliable.

A third distinction concerns what attention is asked to represent. In SEED, DGA uses a grid over a reference box plus predicted offsets so that the network can focus on relevant regions and capture more informative features. In ABSS, the decisive signal is early attention to prompt core tokens. In AGCL, attention is embodied by class-conditional heatmaps that supervise later training. In AG-CAM, attention is part of a class-specific explanation. In UGDA-Net, attention is modulated by channel variance and paired with entropy weighting to focus on ambiguous boundaries. The unifying role is not the identical form of attention, but the use of attention as a ranking, weighting, or refinement signal.

A common misconception is to treat these methods as interchangeable because they all mention “attention-guided” and “seed.” The cited literature does not support that simplification. The seed object, the attention source, and the optimization target differ sharply across point-cloud detection, diffusion sampling, medical-image weak supervision, ViT explanation, and seedling segmentation. What they share is a disciplined attempt to use intermediate model evidence to prioritize candidates that would otherwise be selected by confidence alone or by brute-force search.