The problem the team attacked is one of the most stubborn in computer vision: reconstructing a full 3D human body from a first-person camera. Head-mounted devices see the world from a perspective no studio camera ever captures — steep downward angles, extreme self-occlusion, and constant scale changes as limbs swing toward and away from the lens. When you look down at yourself, your own torso blocks your legs; your hands appear and vanish; body parts blur together in ways that make standard 3D estimation techniques falter. Yet this egocentric viewpoint is exactly what healthcare monitoring, robotics, and augmented reality increasingly demand, because machines need to understand human behavior from the observer’s own visual perspective.

Existing systems typically handle this by regressing the parameters of SMPL-X, a widely used statistical body model that encodes pose, shape, hand articulation, facial expression, jaw and eye orientation in a high-dimensional parameter space. A single dense regression head predicts all these continuous values at once, but the parameters are strongly coupled through the body model’s kinematics, and the model receives little structural guidance about how the components depend on each other. Under occlusion, truncation, and unusual viewpoints, such approaches rely heavily on broad training coverage and often produce implausible results — joints that hyperextend, limbs that drift, or bodies that penetrate themselves.

The EgoTokens team, led by Gulraiz Khan and Maneeha Rani with Kevin Pimbblet and Kenneth Y. Wertheim among the authors, took a radically different route. Instead of regressing continuous numbers, they reformulated the entire reconstruction task as sequence generation — the same paradigm that powers modern chatbots. Each of the 122 SMPL-X parameter dimensions is quantized into one of 1000 discrete bins, and each bin becomes a dedicated token in an extended vocabulary. The model is then trained to autoregressively generate a 122-token parameter sequence from an image-text prompt, treating the human body as a sentence to be written rather than a set of numbers to be fitted.

This tokenization choice is more technically loaded than it first appears. The team deliberately avoided learned codebooks — the vector-quantization approach used by earlier systems such as VQ-VAE and VQ-HPS — because learned codebooks suffer from codebook collapse, where many entries go unused, and from error propagation, where mistakes made during codebook training contaminate the generation stage. Uniform quantization, by contrast, maps each parameter value directly to an interpretable bin with a known numerical meaning, making the process transparent and directly reversible. The maximum quantization error at 1000 bins is just 0.0032 radians (0.18 degrees) for angular parameters and 0.005 meters for translation — precision comparable to the margin of error in multi-camera studio setups.

But the real breakthrough came from a training innovation the researchers call a distance-aware loss. Standard cross-entropy treats every wrong token prediction identically: predicting bin 500 when the true bin was 501 incurs exactly the same penalty as predicting bin 900, even though the former is nearly correct and the latter is catastrophically wrong. The team solved this by introducing a soft-argmax operator that converts the model’s raw logits into a probability distribution over all 1000 bins, then computes the probability-weighted expected value and penalizes the L2 distance from the true parameter. Combined with cross-entropy in a 0.3-to-0.7 weighting, this distance-aware objective slashed reconstruction error dramatically — from 36.2 mm to 28.20 mm PA-MPVPE on the EgoBody benchmark for the InternVL2-8B backbone.

The results are competitive with, and in some cases superior to, far larger specialized pipelines. On EgoBody, the largest egocentric dataset with ground-truth SMPL-X parameters, the fine-tuned InternVL2-8B achieved 29.49 mm PA-MPJPE and 28.20 mm PA-MPVPE, outperforming the recent SMPLest-X-H40+ (33.5 mm PA-MPVPE) despite that model being trained on fourteen additional datasets. Zero-shot evaluation on the EHF studio dataset — never seen during training — yielded 33.85 mm PA-MPVPE, beating OSX (48.7 mm), PyMAF-X (50.2 mm), and Hand4Whole (50.3 mm), and approaching the performance of SMPLer-X-L32 (37.1 mm). The Qwen2.5-VL-7B backbone produced similarly strong numbers, confirming that the tokenization and loss design generalize across model architectures.

Perhaps the most striking capability is language-guided body shape editing. The team built an augmentation pipeline that generates eight semantically controlled shape variants for each training image — belly fat or slim, legs fat or slim, face fat or slim, height tall or short — by manipulating the shape coefficients of the SMPL-X model. Each variant is paired with a natural-language instruction such as “Reconstruct the SMPL-X body parameters with a fat belly” or “Predict the SMPL-X body parameters with thinner legs.” This expanded the dataset from roughly 150,000 to more than 1.3 million instruction-parameter pairs. Anthropometric validation confirmed the edits produce anatomically correct changes: belly enlargement increased hip circumference by 21.7 percent, leg thickening increased thigh circumference by 22.5 percent, and height scaling changed body height by 10 percent, all with statistically significant differences and no cross-talk between unrelated body regions.

At inference time, the model automatically decides whether to perform plain reconstruction or semantic editing based on the instruction in the prompt — no explicit body-part segmentation or post-processing required. The team demonstrated robustness to paraphrasing: instructions like “increase belly size,” “make the person’s abdomen fat,” and “make the person look heavier in the belly region” all map to the same geometric edit. This suggests the model learns genuine semantic understanding between language and body geometry rather than memorizing fixed instruction templates.

The implications ripple outward. A single fine-tuned vision-language model can now serve as a competitive alternative to specialized 3D reconstruction pipelines, while simultaneously offering interactive, language-driven control over the reconstructed body. The framework is model-agnostic, demonstrated across two distinct backbones, and the training cost — three days on a single NVIDIA A40 GPU for the base model, fifteen days for the augmented version — scales linearly with dataset size and can be reduced through multi-GPU distribution. Future work, the authors suggest, could explore adaptive loss weighting, extend quantization to represent temporal information in video, and add further semantic editing directions such as athletic build or broader shoulders. For now, EgoTokens stands as a vivid demonstration that the same models trained to write essays and answer questions can learn to sculpt the human form — one token at a time.

Subject of Research: EgoTokens: Egocentric 3D human reconstruction and language-guided shape editing via quantized position tokens in vision-language models

Article Title: EgoTokens: Egocentric 3D human reconstruction and language-guided shape editing via quantized position tokens in vision-language models

Article References: Khan, G., Rani, M., Ahmed, W., Pimbblet, K., & Wertheim, K. Y. (2026). EgoTokens: Egocentric 3D human reconstruction and language-guided shape editing via quantized position tokens in vision-language models. Machine Learning with Applications, 26, Article 101042. https://doi.org/10.1016/j.mlwa.2026.101042

Image Credits: AI Generated

DOI: 10.1016/j.mlwa.2026.101042

Keywords: EgoTokens, Egocentric, human, reconstruction, language-guided, shape, editing, quantized, position, tokens, vision-language, models

Cardiac MRE fills a diagnostic gap that conventional imaging leaves open. Echocardiography, cardiac CT, and standard cardiac MRI can suggest that the heart muscle has changed — through wall thickness measurements, strain patterns, or T1 mapping values — but none of these tools measures stiffness itself. Instead, they offer indirect surrogates that can be confounded by factors unrelated to the true mechanical properties of the myocardium. Elastography, by contrast, interrogates the tissue directly: an external driver generates low-frequency mechanical vibrations, a passive device on the chest wall transmits those vibrations into the thorax, and phase-contrast MRI captures the shear waves as they propagate through the beating heart. Specialized reconstruction algorithms then convert the wave field into spatially resolved stiffness maps, all without contrast agents or invasive pressure measurements.

The clinical motivation for such a direct measurement is substantial. Increased myocardial stiffness is now recognized as an early and prognostically important feature of a range of cardiac diseases, including myocardial fibrosis, hypertrophic cardiomyopathy, cardiac amyloidosis, restrictive cardiomyopathy, and heart failure with preserved ejection fraction. Detecting these changes before they become irreversible could shape treatment decisions and improve long-term outcomes. Cardiac MRE has already demonstrated the ability to detect elevated myocardial stiffness in cardiac amyloidosis, hypertrophic cardiomyopathy, aortic stenosis, and pulmonary hypertension, and its measurements have been shown to correlate with invasive hemodynamic indices. This track record has led many investigators to regard cardiac MRE as a quantitative imaging biomarker with genuine translational potential.

Yet the technique faces a fundamental physical challenge that distinguishes it from its more established cousin, liver MRE. The liver sits close to the abdominal wall and offers a relatively uniform acoustic pathway for mechanical waves. The heart, by contrast, lies behind the ribs and is partially surrounded by air-filled lung tissue, which attenuates and scatters mechanical vibrations before they can reach the myocardium. The heart also moves rapidly throughout the cardiac cycle and has a complex three-dimensional geometry, both of which complicate the delivery of consistent wave energy to the left ventricular wall. In the single-frequency protocol used in this study, mechanical excitation at one vibration frequency was delivered through a single passive driver, which meant that the driver’s position on the chest wall determined where shear waves entered the thorax and how efficiently they reached the left ventricular myocardium.

Despite this obvious dependence on driver placement, most published cardiac MRE studies have positioned the driver at the apical surface projection — the region to the left of the sternum and just above the xiphoid process — based largely on practical convention rather than on direct comparative evidence. The left chest wall, located at the same axial level as the apical surface projection but along the mid- to anterior-axillary line, represents a clinically feasible alternative when anterior placement is difficult. Such difficulties are not rare in practice: prior thoracic surgery, implanted cardiac devices, breast tissue, obesity, and chest wall deformities can all make anterior placement problematic or impossible. The left chest wall is close to the heart, accessible in the supine position, and compatible with the same fixation band used for conventional placement. The right chest wall, by comparison, sits farther from the heart, and posterior placement is poorly suited to routine supine cardiac MRI workflows. Alternative left lateral placement is therefore a practical and recurring clinical consideration rather than a hypothetical concern.

To compare the two positions objectively, the researchers needed a quantitative measure of wave quality rather than relying on subjective image inspection alone. They turned to the octahedral shear strain signal-to-noise ratio, a metric commonly used in elastography to characterize shear-wave quality. This measure reflects how clearly the propagating waves stand out against background noise in the strain field, and it has been shown to be associated with the reliability of stiffness estimation. When OSS-SNR is low, the inversion algorithms that convert wave data into stiffness values become less robust, and measurement variability tends to increase. A recent normative study of 125 healthy adults had already identified acquisition-related factors — including driver coupling and slice location — as important sources of variability in cardiac MRE, and other investigators had explicitly flagged actuator placement as a key but understudied determinant of wave quality. What was missing was a direct, within-subject comparison of driver positions.

The new study provided exactly that. Twenty-three healthy volunteers with a mean age of 43 years underwent two consecutive cardiac MRE acquisitions at 3.0 Tesla during a single session. The passive driver was placed first at the apical surface projection and then moved to the left chest wall at the same axial level, with the scan repeated after re-localization. Participants were asked to avoid strenuous exercise and stimulant-containing food or drink on the day of scanning, and each rested supine for at least thirty minutes before the examination to limit transient blood pressure swings. Left ventricular myocardial stiffness and OSS-SNR were measured at both positions, and differences were analyzed using paired statistical tests appropriate for within-subject comparisons.

The results were striking. Myocardial stiffness measured at the apical surface projection averaged 9.0 kPa, compared with 14.7 kPa at the left chest wall — a mean difference of 5.8 kPa, which corresponds to a very large effect size. This finding carries an important implication: the two driver positions do not simply differ in image quality, they yield systematically different stiffness values in the same healthy hearts. Any study or clinical program that switches driver positions could introduce a shift in measured stiffness that is large relative to the differences expected between health and disease. Standardizing driver placement is therefore not merely a matter of convenience but a prerequisite for comparing results across studies, institutions, and time points.

The wave quality data help explain why the two positions diverge. OSS-SNR was significantly higher at the apical surface projection than at the left chest wall, again with a substantial effect size. Higher wave quality at the anterior position plausibly reflects a shorter and less obstructed path for mechanical vibrations traveling from the chest surface to the heart, with less attenuation by lung tissue along the way. Better wave delivery in turn supports more reliable stiffness inversion, which is consistent with the reproducibility findings.

Reproducibility was assessed by having observers repeat the stiffness measurements and by comparing measurements between independent observers, using intraclass correlation coefficients and Bland–Altman analysis. At the apical surface projection, stiffness measurements demonstrated excellent agreement both within and between observers, with intraclass correlation coefficients of 0.962 and 0.963 respectively. At the left chest wall, intra-observer agreement remained excellent at 0.914, but inter-observer agreement dropped to 0.839, a level generally classified as good rather than excellent. The pattern suggests that when wave quality degrades, different observers analyzing the same images begin to diverge more in their stiffness estimates, likely because noisier wave fields leave more room for interpretive judgment during region-of-interest placement and quality control.

For clinicians and researchers, the practical takeaway is that the apical surface projection should be considered the preferred driver position for cardiac MRE, offering both higher wave quality and superior reproducibility. The left chest wall remains a viable fallback when anterior placement is impractical, but its use should be documented explicitly, and stiffness values obtained at the two positions should not be interpreted interchangeably. The magnitude of the observed difference — nearly 6 kPa in healthy volunteers — underscores how consequential this seemingly technical choice can be.

Several aspects of the study design merit attention when considering how the findings generalize. The cohort consisted exclusively of healthy volunteers, so the behavior of wave propagation in diseased hearts — with altered geometry, remodeled walls, or pericardial constraints — remains to be established. The sample size of twenty-three, while adequate for a within-subject comparison with large effect sizes, is modest, and the participants spanned a wide age range. Future work will need to determine whether the superiority of the apical position holds in patient populations, across different vibration frequencies, and with the multi-driver or multi-frequency configurations that some cardiac MRE protocols employ. Nevertheless, by directly comparing two clinically realistic driver positions within the same subjects, the study converts a matter of convention into a matter of evidence, and it provides a concrete, actionable standard for anyone seeking to add cardiac MRE to clinical or research practice.

Subject of Research: Cardiac Imaging

Article Title: Evaluating the impact of passive driver position on image consistency and repeatability in 3.0T cardiac magnetic resonance elastography

Article References: Evaluating the impact of passive driver position on image consistency and repeatability in 3.0T cardiac magnetic resonance elastography. (n.d.). https://doi.org/10.3389/fradi.2026.1912855

Image Credits: AI Generated

DOI: 10.3389/fradi.2026.1912855

Keywords: cardiac MRE, myocardial stiffness, reproducibility, apical surface, lateral surface, imaging biomarkers, Evaluating, impact, passive, driver, position, image