HumanSAM: Human-Centric SAM Systems

• HumanSAM is a family of SAM adaptations for human data, integrating prompt-based segmentation and domain priors for tasks like forgery detection and medical imaging.
• It employs human-in-the-loop methods, pseudo-label refinement, and automatic prompt generation to improve annotation efficiency and segmentation accuracy.
• Evaluations across video forgery, surgical annotation, and biomedical imaging demonstrate significant speed and accuracy gains, highlighting its practical impact.

Searching arXiv for HumanSAM and closely related SAM-based medical/human-centric papers to ground the article in current literature. {"query":"HumanSAM arXiv (Liu et al., 26 Jul 2025) HOLa (Schwimmbeck et al., 2024) SAM medical segmentation (Keuth et al., 2024)", "max_results": 10} HumanSAM is a designation used in multiple, distinct research contexts to describe SAM-derived or SAM-inspired systems centered on human data, human interaction, or human-specific analysis. In the most explicit naming sense, “HumanSAM” denotes a framework for classifying human-centric forgery videos into spatial, appearance, motion anomaly, and real categories (Liu et al., 26 Jul 2025). In a broader methodological sense, the term has also been used to characterize human-in-the-loop or human-domain adaptations of the Segment Anything Model (SAM), including HoloLens-guided annotation in surgery (Schwimmbeck et al., 2024), pseudo-label refinement for medical segmentation (Keuth et al., 2024), automatic prompt generation in muscle MRI (Ward et al., 2 Jul 2025), histopathology foundation modeling (Griebel et al., 1 Feb 2025), brain extraction in MRI (Chenna et al., 2024), atlas-assisted fetal brain MRI segmentation (Zeng et al., 22 Jan 2026), and promptable 3D human mesh recovery (Yang et al., 17 Feb 2026). Across these uses, the common thread is the conversion of SAM from a generic promptable segmenter into a practical system for human-centered tasks through prompting, domain adaptation, auxiliary priors, or lightweight human supervision.

1. Terminological scope and definitional variants

The term “HumanSAM” has both a narrow and a broad meaning in the recent literature. In the narrow sense, it refers to the framework introduced in “HumanSAM: Classifying Human-centric Forgery Videos in Human Spatial, Appearance, and Motion Anomaly,” which addresses fine-grained categorization of forged human videos rather than only binary real-versus-fake detection (Liu et al., 26 Jul 2025). In that formulation, the central objective is 4-way classification with labels 0: spatial anomaly, 1: appearance anomaly, 2: motion anomaly, and 3: real video (Liu et al., 26 Jul 2025).

In the broader sense, “HumanSAM” functions as a conceptual umbrella for systems that adapt SAM to human biomedical data, human anatomy, or human-in-the-loop workflows. The supplied literature explicitly frames HOLa as a “HumanSAM” instantiation in which a human provides a single seed point and optional quality control while SAM and SAM-Track perform the pixel-wise work (Schwimmbeck et al., 2024). The semi-supervised pseudo-labeling work “SAM Carries the Burden” likewise describes “HumanSAM” as the problem of turning a powerful, generic SAM model into a practical, high-quality segmentation system in a specific domain using limited labels, pseudo-labels, and possibly humans in the loop (Keuth et al., 2024). ADA-SAM is described as a “HumanSAM-like approach” for human clinical MRI (Ward et al., 2 Jul 2025), and PathoSAM is presented as a SAM-style foundation model for human tissue images in digital pathology (Griebel et al., 1 Feb 2025).

This multiplicity of usage suggests that “HumanSAM” is not yet a single canonical architecture. A plausible implication is that the term now names a family of design patterns organized around promptability, human specificity, and annotation efficiency rather than one fixed model.

2. HumanSAM as human-centric forgery video classification

In its explicit titular usage, HumanSAM is an end-to-end framework for multi-class human-centric forgery video classification (Liu et al., 26 Jul 2025). It was introduced to address the limitation of binary detection, which is characterized as less reliable, not interpretable, and difficult to debug or update when generative models change (Liu et al., 26 Jul 2025). The method instead aims to determine whether a video is real or, if fake, whether it exhibits human spatial anomalies, human appearance anomalies, or human motion anomalies (Liu et al., 26 Jul 2025).

The anomaly taxonomy is central. Human spatial anomaly denotes violations of 3D spatial logic and object interaction, such as hands or rigid objects passing through other objects, with cues derived from depth maps and optical-flow-based warping error between consecutive depth maps (Liu et al., 26 Jul 2025). Human appearance anomaly denotes inconsistencies in identity, shape, or appearance over time, including changes in age, handedness, finger count, or object persistence, quantified by framewise feature similarity using CLIP and DINOv2 with a sliding-window similarity measure (Liu et al., 26 Jul 2025). Human motion anomaly denotes violations of biomechanically plausible motion or temporal consistency, measured via optical-flow warping error on RGB frames (Liu et al., 26 Jul 2025).

The architecture constructs a Human Forgery Representation from two branches (Liu et al., 26 Jul 2025). The video understanding branch uses InternVideo2 and produces pooled semantic and spatiotemporal features, yielding a representation of dimension 2816 after average and attention pooling concatenation (Liu et al., 26 Jul 2025). The spatial depth branch uses a frozen Depth Pro encoder on the first and middle frames, generating features $$F_y \in \mathbb{R}^{2 \times 1024 \times 48 \times 48}$$ and a pooled vector $f_y \in \mathbb{R}^{1024}$ (Liu et al., 26 Jul 2025). A linear layer maps the video features to 1024 dimensions and a learnable scalar $\alpha$ fuses the two branches according to

$$Y = \text{PROJ}\left( \alpha \cdot f_x' + (1 - \alpha) \cdot f_y \right),$$

where $Y \in \mathbb{R}^4$ are pre-softmax logits (Liu et al., 26 Jul 2025).

Training uses a rank-based confidence enhancement strategy derived from three prior anomaly scores $S_s(v_i)$, $S_a(v_i)$, and $S_m(v_i)$, which are converted to ranks $R_s(v_i)$, $R_a(v_i)$, and $f_y \in \mathbb{R}^{1024}$0 (Liu et al., 26 Jul 2025). The dominant anomaly label is assigned by

$f_y \in \mathbb{R}^{1024}$1

Within each anomaly class, a normalized rank $f_y \in \mathbb{R}^{1024}$2 yields a sample weight

$f_y \in \mathbb{R}^{1024}$3

and the weighted loss is

$f_y \in \mathbb{R}^{1024}$4

for base multi-class cross-entropy losses $f_y \in \mathbb{R}^{1024}$5 (Liu et al., 26 Jul 2025).

HumanSAM is trained and evaluated on the HFV dataset, described as the first public benchmark for multi-class human-centric forgery video classification (Liu et al., 26 Jul 2025). Generated content is sourced from nine models—MiniMax, Runway Gen-3, Vchitect-2.0 (VEnhancer), Kling, CogVideoX-5B, Vchitect-2.0-2B, CogVideoX-2B, Pika, and Runway Gen-2—while real videos come from Kinetics-400 (Liu et al., 26 Jul 2025). On multi-class classification across eight evaluation generators, HumanSAM achieves average ACC $f_y \in \mathbb{R}^{1024}$6 and average AUC $f_y \in \mathbb{R}^{1024}$7, compared with TimeSformer at ACC $f_y \in \mathbb{R}^{1024}$8, AUC $f_y \in \mathbb{R}^{1024}$9, and MM-Det at ACC $\alpha$0, AUC $\alpha$1 (Liu et al., 26 Jul 2025). On binary classification, it achieves average ACC $\alpha$2 and average AUC $\alpha$3 (Liu et al., 26 Jul 2025). The ablation study reports ACC/AUC of $\alpha$4 for the InternVideo2 backbone only, $\alpha$5 with concatenated pooled features, $\alpha$6 after adding the depth branch, and $\alpha$7 for the full system with rank-based loss, indicating that the depth branch is a major contributor (Liu et al., 26 Jul 2025).

3. HumanSAM as a human-in-the-loop annotation paradigm

A second major usage of HumanSAM is as a light human-in-the-loop annotation workflow. HOLa, “HoloLens Object Labeling,” is the clearest example of this interpretation (Schwimmbeck et al., 2024). HOLa is a Unity–Python application designed to automatically generate pixel-wise object annotations from Microsoft HoloLens 2 RGB data, targeting the need for large quantities of labeled data in medical AR and surgical navigation (Schwimmbeck et al., 2024). It combines HoloLens 2 as a data acquisition and interaction device with SAM-Track, which builds on SAM, as a zero-shot, sequence-aware annotator (Schwimmbeck et al., 2024).

The interaction model is deliberately minimal. During recording mode, an AR sphere cursor indicates the center of the recorded frame, the user aligns it with the object of interest by moving their head, and a voice command “Start” stores the cursor location as the seed point and begins recording (Schwimmbeck et al., 2024). The user ends recording with “Stop” (Schwimmbeck et al., 2024). In labeling mode, the first-frame center point is passed to SAM as a point prompt, SAM returns three mask proposals and associated internal scores, and HOLa selects

$\alpha$8

before passing the selected mask to a DeAOT-based tracker for propagation through the sequence (Schwimmbeck et al., 2024). The system uses SAM with a ViT-H backbone, disables the default “Segment Everything” mode in SAM-Track, and tracks only one object throughout the sequence (Schwimmbeck et al., 2024).

The human role is therefore reduced to high-level object choice, a single point prompt, and optional post hoc checking (Schwimmbeck et al., 2024). The paper explicitly states that after the initial seed point, everything else is automated: SAM segmentation on the first frame, mask propagation for the full video, and no frame-wise corrections, additional prompts, or steering during tracking (Schwimmbeck et al., 2024). This division of labor is what the paper identifies as the “HumanSAM” perspective (Schwimmbeck et al., 2024).

HOLa is evaluated in five scenarios of increasing complexity: liver phantom, bone phantom, OR phantom, surgery scene 1, and surgery scene 2 (Schwimmbeck et al., 2024). The 90-frame evaluation against a reference annotator reports mean Dice $\alpha$9, $$Y = \text{PROJ}\left( \alpha \cdot f_x' + (1 - \alpha) \cdot f_y \right),$$0, $$Y = \text{PROJ}\left( \alpha \cdot f_x' + (1 - \alpha) \cdot f_y \right),$$1, $$Y = \text{PROJ}\left( \alpha \cdot f_x' + (1 - \alpha) \cdot f_y \right),$$2, and $$Y = \text{PROJ}\left( \alpha \cdot f_x' + (1 - \alpha) \cdot f_y \right),$$3, respectively (Schwimmbeck et al., 2024). Annotation throughput is 5 fps for HOLa versus approximately $$Y = \text{PROJ}\left( \alpha \cdot f_x' + (1 - \alpha) \cdot f_y \right),$$4–$$Y = \text{PROJ}\left( \alpha \cdot f_x' + (1 - \alpha) \cdot f_y \right),$$5 fps for human annotators, corresponding to more than $$Y = \text{PROJ}\left( \alpha \cdot f_x' + (1 - \alpha) \cdot f_y \right),$$6 speed-up (Schwimmbeck et al., 2024). In the 10-frame inter-annotator comparison, HOLa–HA1 agreement is very close to human–human agreement in the phantom experiments and somewhat lower in complex clinical scenes, but still of similar order (Schwimmbeck et al., 2024). The authors conclude that HOLa annotation quality is comparable to human annotators, especially in less ambiguous scenes (Schwimmbeck et al., 2024).

This version of HumanSAM is not about retraining or adapting SAM weights. Instead, it operationalizes human-centered prompting with near-minimal interaction. A plausible implication is that one influential strand of HumanSAM research is workflow design: relocating human effort from dense labeling to prompt specification and quality control.

4. HumanSAM in semi-supervised and self-prompting medical segmentation

Several papers extend the HumanSAM idea from interactive annotation to annotation-efficient medical segmentation. “SAM Carries the Burden” uses SAM as a frozen, prompt-based pseudo-label refiner in a five-step semi-supervised pipeline (Keuth et al., 2024). A U-Net $$Y = \text{PROJ}\left( \alpha \cdot f_x' + (1 - \alpha) \cdot f_y \right),$$7 is first trained on a small labeled set, its predictions on unlabeled images are converted into prompts for SAM, SAM refines those into dense masks, and a second U-Net $$Y = \text{PROJ}\left( \alpha \cdot f_x' + (1 - \alpha) \cdot f_y \right),$$8 is trained on the refined pseudo labels (Keuth et al., 2024). Prompt extraction includes connected component filtering via

$$Y = \text{PROJ}\left( \alpha \cdot f_x' + (1 - \alpha) \cdot f_y \right),$$9

followed by morphological operations, bounding boxes, positive points at the center of mass, and negative points from other classes (Keuth et al., 2024). The best-performing strategy uses a bounding box for the initial SAM call and then a self-refinement step in which SAM’s previous mask prediction is given back as a dense prompt together with sparse positive and negative points (Keuth et al., 2024).

The reported gains are substantial. On paediatric wrist bone segmentation, training with SAM-refined pseudo labels improves Dice from $Y \in \mathbb{R}^4$0 to $Y \in \mathbb{R}^4$1, and on dental radiographs from $Y \in \mathbb{R}^4$2 to $Y \in \mathbb{R}^4$3 (Keuth et al., 2024). The paper reports that this outperforms intensity-based post-processing methods, nnU-Net, and the semi-supervised mean teacher approach (Keuth et al., 2024). It also notes that training on unrefined pseudo labels provides almost no gain, which the authors use to argue that SAM-based refinement, not pseudo-labeling alone, is the critical ingredient (Keuth et al., 2024). In the paper’s own framing, this is exactly a HumanSAM-style pattern: generic SAM plus task-specific prompts and pseudo-label generation (Keuth et al., 2024).

ADA-SAM advances this direction by removing manual prompts entirely in a low-label clinical MRI setting (Ward et al., 2 Jul 2025). The target is segmentation of the vastus lateralis and vastus medialis muscles in thigh MRI (Ward et al., 2 Jul 2025). ADA-SAM uses a ViT-B SAM encoder, LoRA adaptation with rank $Y \in \mathbb{R}^4$4, an auxiliary classification branch that outputs one of four slice-level classes, and Grad-CAM-derived automatic bounding-box prompts for SAM’s prompt encoder (Ward et al., 2 Jul 2025). The segmentation branch uses Dice loss,

$Y \in \mathbb{R}^4$5

the classifier uses focal loss,

$Y \in \mathbb{R}^4$6

and the multitask objective is

$Y \in \mathbb{R}^4$7

(Ward et al., 2 Jul 2025). The distinctive mechanism is “gradient feedback”: segmentation gradients update the shared encoder, which improves the classifier’s Grad-CAMs, which in turn provide better prompts for segmentation (Ward et al., 2 Jul 2025).

On 21 patients and 1365 MRI slices, ADA-SAM reports overall Dice $Y \in \mathbb{R}^4$8, $Y \in \mathbb{R}^4$9, $S_s(v_i)$0, and $S_s(v_i)$1 in the semi-supervised settings with 0, 5, 50, and 100 labeled slices, respectively (Ward et al., 2 Jul 2025). These outperform SAM-Mix, which yields $S_s(v_i)$2, $S_s(v_i)$3, $S_s(v_i)$4, and $S_s(v_i)$5 in the same label regimes (Ward et al., 2 Jul 2025). The paper explicitly positions ADA-SAM as a HumanSAM-like approach because it combines domain-specific adaptation, self-prompting, and annotation efficiency (Ward et al., 2 Jul 2025).

Together, these works show a transition from explicit human prompts to automatic prompt generation. This suggests that a major line of HumanSAM research is not only human-in-the-loop interaction, but also the replacement of expensive human prompts with machine-generated surrogates derived from weak labels, classifiers, or prior models.

5. Domain-specialized HumanSAM variants in biomedical imaging

A further interpretation of HumanSAM is as a domain-specialized SAM family model for human biomedical images. PathoSAM exemplifies this in histopathology (Griebel et al., 1 Feb 2025). It starts from SAM’s image encoder, prompt encoder, and mask decoder, then adds an Automatic Instance Segmentation decoder that predicts foreground probability, distance to closest object center, and distance to closest object boundary, followed by seeded watershed (Griebel et al., 1 Feb 2025). PathoSAM is jointly trained for interactive and automatic instance segmentation and can also be extended with a semantic decoder for PanNuke’s five nucleus classes (Griebel et al., 1 Feb 2025).

The model is trained on six histopathology nucleus instance segmentation datasets—CPM15, CPM17, Lizard, MoNuSeg, PanNuke, PUMA, and TNBC—and evaluated on additional out-of-domain datasets including CryoNuSeg, NuInsSeg, MoNuSAC, LynSec/IHC-TMA, NuClick, GlaS, and the SRSA-Net dataset (Griebel et al., 1 Feb 2025). Mean segmentation accuracy is defined as

$S_s(v_i)$6

over thresholds $S_s(v_i)$7 (Griebel et al., 1 Feb 2025). The paper reports that PathoSAM ViT-L with AIS is the best overall automatic nucleus instance segmentation model across 12 datasets and that, in interactive settings with 7 correction iterations, PathoSAM and μSAM achieve almost perfect mean segmentation accuracy, clearly outperforming SAM (Griebel et al., 1 Feb 2025). In semantic segmentation on PanNuke, PathoSAM is second best overall and does not yet match CellViT (Griebel et al., 1 Feb 2025). The explicit relevance to HumanSAM is that the system is presented as a SAM-style model adapted to human tissue images with both automatic and interactive capabilities (Griebel et al., 1 Feb 2025).

FeTal-SAM provides a different domain-specialization strategy by using atlas-assisted dense prompts for fetal brain MRI (Zeng et al., 22 Jan 2026). It builds on Med-SAM-ViT-b with a frozen image encoder, adds a U-Net-like atlas label encoder and an atlas prompt encoder, and combines dense atlas prompts with averaged bounding-box prompts derived from age-matched atlas registrations (Zeng et al., 22 Jan 2026). The method performs binary, per-structure segmentation slice by slice in axial, coronal, and sagittal views, then fuses the three orientations with STAPLE (Zeng et al., 22 Jan 2026). Loss is the sum of Dice and cross-entropy,

$S_s(v_i)$8

and evaluation uses Dice, Jaccard, HD95, and MSD (Zeng et al., 22 Jan 2026).

On the dHCP fetal brain dataset, FeTal-SAM reports average Dice $S_s(v_i)$9, Jaccard $S_a(v_i)$0, HD95 $S_a(v_i)$1 mm, and MSD $S_a(v_i)$2 mm, compared with Med-SAM at Dice $S_a(v_i)$3 and Med-SAM-FT at Dice $S_a(v_i)$4 (Zeng et al., 22 Jan 2026). On the CRL dataset, it reports Dice $S_a(v_i)$5, Jaccard $S_a(v_i)$6, HD95 $S_a(v_i)$7 mm, and MSD $S_a(v_i)$8 mm (Zeng et al., 22 Jan 2026). The paper emphasizes that FeTal-SAM is flexible with respect to label definition because it segments one structure at a time from atlas prompts rather than being trained for one fixed label set (Zeng et al., 22 Jan 2026). It also highlights interpretability: because priors are explicit in the prompts, one can distinguish segmentation driven by image contrast from segmentation driven by spatial priors (Zeng et al., 22 Jan 2026).

The paper “Segment anything model (SAM) for brain extraction in fMRI studies” represents a simpler zero-shot biomedical use case (Chenna et al., 2024). SAM ViT-H/16 is applied slice-by-slice to human MRI volumes rescaled or padded to $S_a(v_i)$9, with no fine-tuning or MRI-specific training (Chenna et al., 2024). Evaluation against manual masks yields IoU $S_m(v_i)$0 for sagittal slices, $S_m(v_i)$1 for coronal slices, and $S_m(v_i)$2 for axial slices across 30 image slices from four subjects (Chenna et al., 2024). Although the study is limited in scale, it illustrates a foundational HumanSAM question: how far a generic SAM can go on human medical images without domain retraining (Chenna et al., 2024).

6. Promptable human modeling beyond segmentation

HumanSAM-like ideas also extend beyond segmentation in the conventional sense. “SAM 3D Body: Robust Full-Body Human Mesh Recovery” introduces a promptable system for single-image full-body 3D human mesh recovery that is explicitly designed in the spirit of the SAM family (Yang et al., 17 Feb 2026). The model predicts body, hand, and foot pose using the new Momentum Human Rig, which decouples skeletal structure and surface shape (Yang et al., 17 Feb 2026). It accepts auxiliary prompts, including 2D keypoints and instance masks, which are integrated into a transformer decoder that cross-attends to image features (Yang et al., 17 Feb 2026).

The architecture uses an image encoder on a $S_m(v_i)$3 human crop, prompt tokens from 2D keypoints, mask embeddings added to image features, and a set of query tokens including an MHR plus camera token, learned 2D and 3D keypoint tokens, and hand position tokens (Yang et al., 17 Feb 2026). The training objective is a weighted sum of 2D keypoint, 3D keypoint, parameter regression, joint-limit, and hand-box losses (Yang et al., 17 Feb 2026). MHR parameter updates are supervised through regressions such as

$S_m(v_i)$4

(Yang et al., 17 Feb 2026).

Prompting has measurable effects. With 3DB-H on COCO and EMDB, adding one correct keypoint prompt improves COCO [email protected] from 86.7 to 90.2 and reduces EMDB MPJPE from 63.3 mm to 60.1 mm; two keypoint prompts raise COCO [email protected] to 93.0 and reduce EMDB MPJPE to 58.9 mm (Yang et al., 17 Feb 2026). Mask conditioning is especially beneficial in multi-person scenes: on Hi4D, PVE drops from 91.4 to 58.3 and MPJPE from 76.4 mm to 47.0 mm when mask prompts are used (Yang et al., 17 Feb 2026). The model is therefore “HumanSAM” in a strong interface sense: prompts allow user-guided disambiguation and correction of full-body 3D reconstructions (Yang et al., 17 Feb 2026).

This broadens the concept substantially. HumanSAM need not denote only segmentation of biomedical structures or videos; it can also denote promptable human-centric modeling systems where SAM-like interaction is transferred to 3D human recovery.

7. Conceptual themes, limitations, and misconceptions

Across these papers, several recurring themes define the HumanSAM landscape. First is promptability. Whether the prompt is a HoloLens seed point (Schwimmbeck et al., 2024), a bounding box and positive/negative points derived from weak predictions (Keuth et al., 2024), a Grad-CAM-generated box (Ward et al., 2 Jul 2025), an atlas-derived dense spatial prior (Zeng et al., 22 Jan 2026), a pathology click or box (Griebel et al., 1 Feb 2025), or 2D keypoints and masks for 3D body recovery (Yang et al., 17 Feb 2026), the unifying mechanism is conditional inference rather than monolithic end-to-end segmentation.

Second is annotation efficiency. HOLa reports more than $S_m(v_i)$5 labeling speed-up (Schwimmbeck et al., 2024); pseudo-label refinement reduces dependence on large labeled datasets (Keuth et al., 2024); ADA-SAM is strongest in 0–50 labeled slice regimes (Ward et al., 2 Jul 2025); FeTal-SAM avoids retraining when label definitions change (Zeng et al., 22 Jan 2026). This suggests that HumanSAM is best understood as a response to the cost structure of dense human annotation.

Third is explicit handling of priors. HumanSAM in forgery analysis fuses semantics, motion, and depth priors (Liu et al., 26 Jul 2025). FeTal-SAM externalizes spatial priors through registered atlases (Zeng et al., 22 Jan 2026). ADA-SAM externalizes localization via classifier-induced CAMs (Ward et al., 2 Jul 2025). These systems do not merely fine-tune SAM; they augment it with additional, structured sources of information.

Several limitations recur as well. HOLa is single-object and sensitive to initialization quality and low-contrast, embedded structures (Schwimmbeck et al., 2024). The pseudo-label refinement pipeline in “SAM Carries the Burden” assumes at most one instance per class per image because it keeps only one connected component $S_m(v_i)$6 per class (Keuth et al., 2024). ADA-SAM depends on CAM quality and has no explicit segmentation pseudo-labeling or consistency regularization (Ward et al., 2 Jul 2025). FeTal-SAM is constrained by atlas quality and registration accuracy, and underperforms 3D baselines on subtle, low-contrast fetal structures (Zeng et al., 22 Jan 2026). PathoSAM’s automatic mode struggles on CryoNuSeg and its semantic segmentation does not yet exceed CellViT (Griebel et al., 1 Feb 2025). HumanSAM for video forgery detection assigns only a single dominant anomaly label even when multiple anomaly types coexist (Liu et al., 26 Jul 2025). SAM 3D Body remains per-person and does not explicitly model human–object or inter-person physical constraints (Yang et al., 17 Feb 2026).

A common misconception would be to treat HumanSAM as a single, standardized model. The literature instead supports a more plural definition. HumanSAM names at least three overlapping ideas: a specific forgery-classification framework (Liu et al., 26 Jul 2025); a family of human-in-the-loop SAM workflows (Schwimmbeck et al., 2024); and a broader class of human-domain SAM adaptations across medical imaging, pathology, and human pose or mesh recovery (Keuth et al., 2024, Ward et al., 2 Jul 2025, Griebel et al., 1 Feb 2025, Zeng et al., 22 Jan 2026, Yang et al., 17 Feb 2026).

Taken together, these works indicate that HumanSAM is becoming a research idiom for adapting promptable foundation models to human-centered tasks. This suggests that future development is likely to continue along three axes already visible in the literature: lighter human supervision, stronger automatic prompt generation, and tighter integration of domain priors with prompt-conditioned foundation models.