SeedPrints: Cross-Domain Machine Signatures
- SeedPrints is a cross-domain concept that denotes machine-readable signatures extracted from heterogeneous sources such as seed images, LLM biases, and 3D model generation outputs.
- In botanical image analysis, SeedPrints pipelines use controlled imaging and blue-background segmentation to derive morpho-colorimetric fingerprints for rapid, reproducible classification.
- In machine learning and digital fabrication, SeedPrints enable intrinsic model attribution and prompt-conditioned 3D printing workflows by capturing persistent, verifiable system biases.
SeedPrints is used in the supplied research literature as a designation for machine-readable signatures extracted from heterogeneous objects and models. In botanical image analysis, it denotes or directly motivates pipelines that transform seed images into segmented objects, morpho-colorimetric descriptors, and classification outputs. In LLM provenance, "SeedPrints" names a passive fingerprinting method that attributes lineage from initialization-seed-dependent biases present before training. In digital fabrication, the same term appears as a systems concept for prompt-conditioned generation of directly printable 3D models. Across these settings, the shared idea is not a single modality but a persistent signature: a compact representation that supports identification, comparison, or verification under controlled acquisition or probing conditions (Vale et al., 2020, Loddo et al., 2021, Sarker et al., 2024, Tong et al., 30 Sep 2025, Wu et al., 20 Nov 2025).
1. Terminological scope
In the botanical papers, SeedPrints is not the title of a standalone method, but the term is used naturally for a pipeline in which scanner-based seed images are segmented and converted into quantitative fingerprints built from morphology, texture, and colour. The segmentation work in "A new automatic approach to seed image analysis: From acquisition to segmentation" provides what the paper description explicitly calls the “front end” of such a pipeline, while "An effective and friendly tool for seed image analysis" provides the corresponding feature-extraction and classification framework inside ImageJ (Vale et al., 2020, Loddo et al., 2021).
In the machine-learning literature, "SeedPrints: Fingerprints Can Even Tell Which Seed Your LLM Was Trained From" defines SeedPrints as an intrinsic fingerprinting method for LLM provenance. Here the “seed” is the random initialization seed, and the fingerprint is a seed-dependent bias in token or hidden-state preferences that exists already at initialization and persists throughout training (Tong et al., 30 Sep 2025).
A further usage appears in "From Prompts to Printable Models: Support-Effective 3D Generation via Offset Direct Preference Optimization", where SeedPrints-style systems are described as workflows in which a user enters a prompt and receives a directly 3D-printable, support-efficient model. In that setting the term functions as a conceptual label for fabrication-aware generative pipelines rather than for a named algorithm (Wu et al., 20 Nov 2025).
This suggests that SeedPrints is best understood as a cross-domain label for persistent, technically exploitable signatures. The exact carrier of the signature differs—seed morphology, detector outputs, initialization-induced biases, or support-efficient geometry—but the operational goal remains attribution, recognition, or reproducible downstream use.
2. Scanner-based seed segmentation as the front end of botanical SeedPrints
The segmentation paper introduces a single-image, fully automatic seed-segmentation method implemented as an open-source Java plugin for ImageJ, "blueBackgroundSeedsSegmenter". It was tested on 3,386 seed samples from 120 Fabaceae species, using flatbed-scanner acquisition and four different blue background hues. The method replaces the authors’ earlier Double Image Method, in which each sample had to be scanned twice on white and black backgrounds and then manually segmented; that older workflow was time-consuming and sensitive to seed displacement between scans (Vale et al., 2020).
The acquisition protocol is highly controlled. Images were acquired with an Epson Perfection V550 flatbed scanner at a minimum of 400 dpi; timing tests used images of size pixels in JPEG format. Seeds are placed separately on the scanner glass, avoiding overlaps, and a tray lined with blue paper provides a uniform blue background. The paper reports robustness across four blue tones, ranging from light to dark blue, rather than dependence on a single calibrated shade (Vale et al., 2020).
Algorithmically, the method combines RGB and HSB colour information with simple morphological post-processing. The core step is blue-background suppression: pixels are classified as background when they exhibit blue-dominant intensity and blue-range hue characteristics. After this suppression step, the plugin performs region filling to remove holes inside seed regions and removes small artefacts by discarding connected components whose area is less than one third of the largest object’s area. The output is a clean binary mask suitable for measurement and batch processing (Vale et al., 2020).
The reported performance is central to its role in SeedPrints-style pipelines. For the Single Image Method, the comparison set consisted of 480 images with 13,544 seeds; the Double Image Method comparison used 240 images with 6,772 seeds. The new plugin segmented all 480 images without generating object detection errors, while the earlier method produced segmentation errors in 26 samples because of scan mis-registration. Average segmentation time was $0.02$ s per image for the new plugin versus $63$ s for the manual method, and acquisition time was approximately $59.9$ s per sample for the single-scan workflow versus $96.5$ s for the double-scan workflow (Vale et al., 2020).
Within a SeedPrints interpretation, this segmentation stage is the infrastructural prerequisite for any later morphometric or colour-based fingerprint. The paper’s significance lies less in taxonomic modeling than in establishing a reproducible foreground–background separation regime that scales to large image collections and heterogeneous seed shapes.
3. Feature-vector SeedPrints in ImageJ
The ImageJ framework paper operationalizes botanical SeedPrints as explicit per-seed feature vectors. It introduces two plugins: SeedsAnalyser, which extracts descriptors from seed images, and SeedsClassifier, which trains and applies machine-learning classifiers through Weka. The framework works from RGB images acquired on a quasi-uniform blue background and generates up to 64 handcrafted descriptors per seed: 32 morphological, 16 texture, and 16 colour features (Loddo et al., 2021).
Segmentation in SeedsAnalyser is based on the blue channel. The plugin extracts the blue channel from an RGB image, applies an automatic thresholding procedure to generate a binary mask, and then uses ParticleAnalyzer-style connected-component logic with user-defined area and optional circularity constraints. Accepted components are treated as individual seed ROIs, cropped, and passed to the feature-extraction stage (Loddo et al., 2021).
The descriptor set is extensive. Morphological features include area, perimeter, Feret, Breadth, aspect ratio, convex area, convex perimeter, equivalent diameter, solidity, convexity, concavity, rectangularity, sphericity, elongation, roundness, thinness ratio, jaggedness, circularity, and asymmetry measures such as FBtoCM. Texture features are computed on grayscale seed regions and include first-order statistics such as mean, standard deviation, skewness, kurtosis, entropy, and uniformity, together with Haralick GLCM descriptors—energy, contrast, correlation, and homogeneity—computed in four directions. Colour features are derived from RGB and HSV spaces and include channel means, channel standard deviations, square roots of channel means, mean RGB, and mean and standard deviation for hue, saturation, and value (Loddo et al., 2021).
These vectors are the seed prints in the strongest literal sense: compact, object-wise numerical signatures intended for storage, comparison, and classification. The classification plugin accepts ARFF feature tables and exposes four Weka models: k-NN, Naive Bayes, Random Forest, and SVM. Training uses 10-fold cross-validation, and the model with the largest ROC AUC is selected (Loddo et al., 2021).
The classification experiments used 1,988 seeds from 23 Fabaceae families, selected from a larger database of 3,386 samples belonging to 120 plant species. Results show that handcrafted SeedPrints can be highly effective under controlled imaging. Random Forest with colour features achieved accuracy, specificity, sensitivity, macro average geometric, mean F-measure, and $0.02$0 macro average arithmetic. Random Forest with all features achieved $0.02$1 accuracy. The best CNN baseline, SeedNet, reached $0.02$2 accuracy but required 12 minutes of training, whereas the best handcrafted-feature Random Forest required 29 seconds (Loddo et al., 2021).
The paper’s position is therefore not that handcrafted SeedPrints dominate deep learning, but that they offer a complete, open-source, ImageJ-integrated workflow with comparatively low computational cost and strong performance in fine-grained seed classification under controlled acquisition.
4. Deep visual SeedPrints for cannabis seed variant detection
The cannabis study extends the SeedPrints idea from feature-table classification to instance-level deep detection. Its task is object detection and classification of cannabis seed variants: given an RGB photograph of seeds on a white background, the model predicts a bounding box for each seed and assigns one of 17 class labels. The dataset contains 3,335 original images at $0.02$3 resolution, reduced to 3,319 after removal of 16 blurry images; all were captured with an Apple iPhone 13 Pro under a consistent white background but varying lighting conditions and angles (Sarker et al., 2024).
Annotation is partially automated. The authors use Grounding DINO, an open-set object detector, to generate seed bounding boxes rather than drawing all annotations manually. The final split is 1,771 training images, 723 validation images, and 825 test images. Albumentations is used for geometric, photometric, and blur augmentations, including random flips, shifts, scaling, rotations up to $0.02$4, brightness/contrast changes, RGB shifts, hue–saturation–value perturbations, channel shuffle, and blur operations (Sarker et al., 2024).
The detector is Faster R-CNN with a ResNet-50 backbone and Feature Pyramid Network, pretrained on COCO and implemented in PyTorch and MMDetection on an NVIDIA RTX 3090 GPU. The study compares six variants that differ in augmentation and in the bounding-box regression loss: mBaseline, mL1, mIoU, mGIoU, mDIoU, and mCIoU. Training runs for 100 epochs with SGD, momentum $0.02$5, weight decay $0.02$6, image resizing to $0.02$7, and batch sizes of 2 for training and 1 for validation and testing (Sarker et al., 2024).
The best overall model is mL1, which combines data augmentation with L1 loss for box regression. On the test set it achieves $0.02$8 mAP@$0.02$9, $63$0 mAP@$63$1, average recall $63$2, F1 $63$3, inference time $63$4 ms, and $63$5 FPS. mIoU provides the best speed–accuracy tradeoff, with $63$6 mAP@$63$7 and $63$8 ms inference time. The non-augmented baseline is materially worse, at $63$9 mAP@$59.9$0 and F1 $59.9$1 (Sarker et al., 2024).
Class-wise performance indicates that the learned visual fingerprints are strong for most varieties but not uniformly so. For mL1 under mAP@$59.9$2, HKRKU and SDA reach $59.9$3, PD reaches $59.9$4, TFT reaches $59.9$5, and KKV reaches $59.9$6; by contrast, CP is the most difficult class at $59.9$7. This pattern supports the paper’s interpretation that most varieties are visually distinctive in RGB imagery, while a small subset is visually ambiguous or underrepresented (Sarker et al., 2024).
Relative to the handcrafted ImageJ pipeline, this work shifts SeedPrints from explicit morpho-colour descriptors to learned detector representations. The emphasis is on non-destructive screening, quality control, and seed-lot verification rather than on manually interpretable feature tables.
5. SeedPrints as intrinsic fingerprints of LLMs
The 2025 SeedPrints paper introduces a different meaning of the term: a passive LLM fingerprinting method for provenance verification and model attribution. Its central claim is that random initialization leaves seed-dependent token-selection biases that are already measurable in untrained models and remain detectable throughout training. The paper characterizes this as a stronger, more intrinsic notion of fingerprinting than post-hoc methods based on training dynamics, corpus exposure, hyperparameters, or textual style (Tong et al., 30 Sep 2025).
The method begins by probing a model $59.9$8 with many random inputs $59.9$9, often pseudo-token sequences constructed from random embeddings. Let $96.5$0 be the mean output vector over those inputs. SeedPrints defines the identity index set as the indices of the $96.5$1 smallest entries of $96.5$2: $96.5$3 The intuition is that low-mean output dimensions are more stable under training and therefore preserve seed-specific structure more reliably than strongly activated ones (Tong et al., 30 Sep 2025).
For a source model $96.5$4 and suspect model $96.5$5, the method intersects the identity sets, applies a row-wise softmax to the restricted outputs on the common indices, and computes a Kendall–Tau rank correlation for each shared identity index across the random input ensemble. Those correlations are compared with a null distribution constructed from independent Gaussian random matrices, and a one-sided $96.5$6-test or Mann–Whitney $96.5$7-test is used with significance level $96.5$8. The decision rule is a same-lineage verdict when the $96.5$9-value is below 0 (Tong et al., 30 Sep 2025).
The experiments span LLaMA-style and Qwen-style models trained from scratch, checkpoints across training, continued pretraining on new domains, large pretrained derivatives, and the LeafBench benchmark. For randomly initialized models, cross-seed pairs have 1, while same-seed initialization-to-trained comparisons produce 2-values from 3 down to 4, indicating persistence of the fingerprint across training. The paper also reports that identical training data and order do not cause fingerprints from different seeds to converge (Tong et al., 30 Sep 2025).
The domain-shift results are especially important. Starting from an OpenWebText base model, the authors continue training on TinyStoriesV2 and on The Stack. True descendants from seed 1000 yield 5-values that are often effectively zero, whereas distractor models trained on the same new corpus but from a different seed yield 6-values around 7 to 8. On large pretrained derivatives of Llama-2-7B, including finance, Vicuna, WizardMath, Meditron, and CodeLlama variants, SeedPrints still reports hidden-state-correlation 9-values below 0 (Tong et al., 30 Sep 2025).
On LeafBench, which includes 696 model pairs, 58 models, and six parameter-altering transformations—Instruct, Finetune, PEFT, Quantization, Merge, and Distillation—SeedPrints remains competitive with strong weight-based methods and clearly surpasses behavior-based baselines under distribution shift. When 1-values are converted to scores 2, the paper reports overall AUC 3 and KS statistic 4; by transformation type, SeedPrints reports overall average AUC 5 and KS 6, with AUC 7 on Instruct, PEFT, Quantization, Merge, and Distillation (Tong et al., 30 Sep 2025).
The conceptual novelty is that the fingerprint is framed as “Galtonian”: unique, stable from birth, and persistent for the model’s entire lifecycle. In contrast to botanical SeedPrints, which summarize observable phenotype, LLM SeedPrints summarize initialization-induced latent bias structure.
6. SeedPrints-style support-aware 3D generation
In the digital-fabrication literature, SeedPrints appears as a systems idea for going from prompts to printable models. The relevant framework is SEG, which fine-tunes the TRELLIS text-to-3D generator using Offset Direct Preference Optimization so that generated meshes require less support material under 3D printing. The paper explicitly describes SeedPrints-style systems as ones in which a user types a prompt and receives a directly 3D-printable, support-efficient model (Wu et al., 20 Nov 2025).
The fabrication criterion is based on the standard self-support condition for FDM printing. Given print direction 8, face normal 9, and maximal self-support angle 0, a face is considered self-supporting if
1
Faces that violate this condition are treated as risky overhangs. SEG estimates support volume geometrically using AABB-based projection and tetrahedral approximations accelerated with Embree, then normalizes by object volume to obtain the normalized support volume
2
Lower NSV corresponds to greater support efficiency (Wu et al., 20 Nov 2025).
Training data are constructed from printable meshes in Thingi10k. After manifold cleanup, each model is rendered from eight views and captioned with Cap3D, yielding approximately 9,840 prompt–model pairs. For each prompt, TRELLIS samples 10 latent codes, support simulation assigns NSV values to the resulting meshes, and pairwise preferences are formed so that lower-NSV samples are preferred. Fine-tuning uses LoRA with rank 3 and alpha 4, ODPO with offset weight 5, Adam, learning rate 6, batch size 7 with gradient accumulation, 50,000 steps, and 8 NVIDIA A100 GPUs with 40 GB each (Wu et al., 20 Nov 2025).
The quantitative results show substantial support-volume reduction. On Thingi10k-Val, TRELLIS reports NSV 9, DPO 0, DRO 1, and SEG 2; NSV3 is 4, 5, 6, and 7, respectively, while SEG reaches SEC 8. On GPT-3DP-Val, TRELLIS reports NSV 9, DPO 0, DRO 1, and SEG 2, with SEG reaching SEC 3. Ablations show that removing the offset reduces SEG to DPO-level performance, while increasing the offset weight to 4 harms results (Wu et al., 20 Nov 2025).
This is not SeedPrints in the narrow LLM-provenance sense. However, the paper shows how the term can be extended to a fabrication-aware workflow in which a prompt-conditioned model is steered toward outputs that carry a usable physical signature: printability under a fixed manufacturing regime.
7. Comparative interpretation, common misconceptions, and limitations
A common misconception would be to treat SeedPrints as a single fixed method. In the supplied literature it refers to at least three distinct constructs: morpho-colorimetric seed fingerprints in plant image analysis, seed-dependent intrinsic fingerprints for LLM provenance, and a broader prompt-to-print systems concept in 3D generation (Loddo et al., 2021, Tong et al., 30 Sep 2025, Wu et al., 20 Nov 2025). Another misconception would be to assume that all SeedPrints are post-hoc summaries. The LLM work is explicitly the opposite: the fingerprint exists at initialization and is intended to support “birth-to-lifecycle” verification (Tong et al., 30 Sep 2025).
The main limitations differ by domain. Botanical SeedPrints depend on controlled acquisition: blue, quasi-uniform backgrounds, non-overlapping seeds, scanner-based imaging, and in the segmentation paper, validation on Fabaceae seeds. The method does not attempt touching-seed separation and may require threshold adjustment for blue-tinted, highly reflective, or scanner-shifted conditions (Vale et al., 2020). The ImageJ feature-extraction pipeline similarly assumes well-separated seeds and does not report explicit feature selection or extensive classifier hyperparameter optimization (Loddo et al., 2021).
The cannabis detector is constrained by its data regime: one camera model, a white background, locally sourced Thai varieties, and a finite dataset of 3,319 images. The low mAP for CP indicates that purely visual RGB SeedPrints are not equally separable across all varieties, and the authors note that industrial-grade certification may require additional modalities for legally critical decisions (Sarker et al., 2024).
LLM SeedPrints require stronger access assumptions. The method assumes availability of a candidate source model and access to logits or final hidden states under chosen random inputs. It does not provide formal finite-sample error bounds, and adversarial training specifically aimed at erasing identity indices is left open (Tong et al., 30 Sep 2025). The 3D-generation setting has its own restrictions: manifold-mesh assumptions, fixed upright orientation, an approximate support simulator rather than an exact slicer, and a target regime centered on FDM with 5 (Wu et al., 20 Nov 2025).
Taken together, these works indicate that SeedPrints is best read as a family of signature-construction strategies rather than a single algorithm. This suggests a unifying abstraction: a SeedPrint is a representation designed to preserve identity-relevant structure under nuisance variation while remaining operational for segmentation, classification, attribution, or fabrication-aware generation.