arxiv: 2604.21681 · v1 · submitted 2026-04-23 · 💻 cs.CV

Recognition: unknown

Sapiens2

Rawal Khirodkar , He Wen , Julieta Martinez , Yuan Dong , Su Zhaoen , Shunsuke Saito

Authors on Pith no claims yet

Pith reviewed 2026-05-09 21:56 UTC · model grok-4.3

classification 💻 cs.CV

keywords human-centric visiontransformer modelspose estimationbody-part segmentationnormal estimationmasked reconstructioncontrastive pretraininghigh-resolution models

0 comments

The pith

Sapiens2 uses combined masked reconstruction and contrastive pretraining on one billion human images to set new benchmarks on pose estimation, body-part segmentation, and normal prediction while adding pointmap and albedo tasks.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents Sapiens2, a family of high-resolution transformer models for human-centric vision, that merges masked image reconstruction with self-distilled contrastive pretraining to learn both fine details and broad semantics. This approach, trained on a curated set of one billion high-quality human images and supported by architectural updates for stability and longer schedules, produces measurable gains over the prior version. A reader would care because the method shows how a single pretraining recipe can improve dense prediction tasks such as pose, segmentation, and surface normals while opening new output capabilities like point maps and albedo without task-specific redesigns.

Core claim

Sapiens2 improves over its predecessor by combining masked image reconstruction with self-distilled contrastive objectives during pretraining on one billion curated human images, incorporating architectural advances from frontier models that support stable longer training, and adopting windowed attention in hierarchical variants for 4K resolution; the resulting models raise performance by 4 mAP on pose estimation, 24.3 mIoU on body-part segmentation, and 45.6 percent lower angular error on normal estimation while extending to pointmap and albedo estimation.

What carries the argument

The unified pretraining objective that pairs masked image reconstruction with self-distilled contrastive learning on a 1-billion-image human dataset, together with hierarchical transformers that use windowed attention for high-resolution outputs up to 4K.

If this is right

Pose estimation accuracy rises by 4 mAP points over the prior generation.
Body-part segmentation quality increases by 24.3 mIoU.
Surface normal estimation error falls by 45.6 percent in angular measure.
The same models gain the ability to output pointmaps and albedo maps without separate training.
Hierarchical variants with windowed attention maintain stability at 4K output resolution after pretraining at 2K.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The results imply that a single pretraining mixture can reduce the need for separate low-level and high-level feature learners in human vision pipelines.
Similar unified objectives might transfer to other dense prediction domains if large curated datasets become available.
The stability gains from the architectural updates could support even longer training runs or larger parameter counts without additional regularization.

Load-bearing premise

The reported performance gains arise mainly from the specific mix of pretraining objectives, data volume and curation, and architectural tweaks rather than from hidden differences in data quality or simple increases in model scale.

What would settle it

Retraining an otherwise identical model using only one of the two pretraining objectives on the same 1B-image set and checking whether the drops in pose mAP, segmentation mIoU, and normal error match the full reported improvements.

Figures

Figures reproduced from arXiv: 2604.21681 by He Wen, Julieta Martinez, Rawal Khirodkar, Shunsuke Saito, Su Zhaoen, Yuan Dong.

**Figure 1.** Figure 1: SAPIENS2 for dense-prediction tasks. We compare 1B models from both generations on segmentation, depth, and normals. Sapiens2 improves over Sapiens with stronger generalization and sharper segmentation of rare classes (lips, tongue, earrings), achieving pixel-accurate hair segmentation. On geometric tasks (depth, normals), it captures subtler facial, clothing, and hair details—all without task-specific a… view at source ↗

**Figure 2.** Figure 2: k-NN comparison using [CLS] token. SAPIENS2 learns a more discriminative, human-semantic feature space—grouping visually similar concepts and improving retrieval performance at high resolution. We extensively evaluate SAPIENS2 across various tasks and benchmarks [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Figure 3: Human-centric attention. Visualization of [CLS]-token self-attention across heads in the final layer. information into the learned features. LiftedCL (Chen et al., 2022) incorporates an adversarial loss to supervise the lifted 3D skeletons, explicitly embedding 3D human structure information for human-centric pretraining. SapiensID (Kim et al., 2025b) trains a model specifically for person re-identificatio… view at source ↗

**Figure 4.** Figure 4: SAPIENS2 Pretraining. We combine the masked reconstruction loss (Lmae) with a global contrastive loss on [CLS] (Lcl). Multiple image views are generated, and a student–teacher framework matches predicted distributions across views. Lmae helps the model learn low-level details (e.g.texture) for high-fidelity dense tasks, while Lcl improves semantic understanding across human images. the loss averages MSE ov… view at source ↗

**Figure 5.** Figure 5: Windowed self-attention for 4K resolution. We revise the backbone to stably scale to 5B parameters, increase the input resolution from 1K to 4K, and maintain compatibility with sparse masked pretraining. The mid-depth blocks use grouped-query attention (GQA) (Ainslie et al., 2023), while the early and late blocks use standard multi-head self-attention. We replace the feed-forward layers with gated SwiGLU… view at source ↗

**Figure 6.** Figure 6: Post-Training Annotations. We annotated 100K in-the-wild images with pose (a) and segmentation (b), class vocabulary is also extended to include eyeglasses (in cyan). For pointmap, normal, albedo (c), we improve our synthetic assets to capture finer geometric details and color variations. 5 POST-TRAINING We fine-tune the pretrained backbone on five human-centric tasks—pose estimation, body-part segmentati… view at source ↗

**Figure 7.** Figure 7: Body-part segmentation using our 1B-4K model. 8 [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗

**Figure 8.** Figure 8: (Top) Pointmap qualitative comparison of Sapiens2-1B with MoGe (Wang et al., 2025b). (Bottom) Depth visualized from the predicted pointmap, along with surface normals and novel 3D viewpoints. 9 [PITH_FULL_IMAGE:figures/full_fig_p009_8.png] view at source ↗

**Figure 9.** Figure 9: Normal prediction. Qualitative comparison of Sapiens2-1B with DAViD (Saleh et al., 2025) [PITH_FULL_IMAGE:figures/full_fig_p010_9.png] view at source ↗

**Figure 10.** Figure 10: Albedo estimation using Sapiens2-1B. Our model effectively encodes low-level details crucial for albedo estimation and generalizes well to in-the-wild images, despite being trained on limited synthetic data. 10 [PITH_FULL_IMAGE:figures/full_fig_p010_10.png] view at source ↗

**Figure 11.** Figure 11: We randomly mix blockwise and patchwise masking to provide coarse occlusions. For MAE pre [PITH_FULL_IMAGE:figures/full_fig_p016_11.png] view at source ↗

**Figure 12.** Figure 12: We visualize the encoder features using PCA (3 major components) with different colors. We use [PITH_FULL_IMAGE:figures/full_fig_p017_12.png] view at source ↗

**Figure 13.** Figure 13: In addition to in-the-wild annotations we also use capture-studio 3D triangulated ground-truth 308 [PITH_FULL_IMAGE:figures/full_fig_p018_13.png] view at source ↗

**Figure 14.** Figure 14: Top-down 308 keypoint predictions using Sapiens2-1B model on in-the-wild images. [PITH_FULL_IMAGE:figures/full_fig_p019_14.png] view at source ↗

**Figure 15.** Figure 15: Body-part segmentation (29 classes) using Sapiens2-1B on real-world images. [PITH_FULL_IMAGE:figures/full_fig_p020_15.png] view at source ↗

**Figure 16.** Figure 16: Pointmap using Sapiens2-1B. For each image, we visualize the absolute depth derived from the [PITH_FULL_IMAGE:figures/full_fig_p021_16.png] view at source ↗

**Figure 17.** Figure 17: Pointmap using Sapiens2-1B. For each image, we visualize the absolute depth derived from the [PITH_FULL_IMAGE:figures/full_fig_p022_17.png] view at source ↗

**Figure 18.** Figure 18: Surface normal prediction using Sapiens2-1B. [PITH_FULL_IMAGE:figures/full_fig_p023_18.png] view at source ↗

**Figure 19.** Figure 19: Albedo (base color) prediction using Sapiens2-1B at [PITH_FULL_IMAGE:figures/full_fig_p024_19.png] view at source ↗

read the original abstract

We present Sapiens2, a model family of high-resolution transformers for human-centric vision focused on generalization, versatility, and high-fidelity outputs. Our model sizes range from 0.4 to 5 billion parameters, with native 1K resolution and hierarchical variants that support 4K. Sapiens2 substantially improves over its predecessor in both pretraining and post-training. First, to learn features that capture low-level details (for dense prediction) and high-level semantics (for zero-shot or few-label settings), we combine masked image reconstruction with self-distilled contrastive objectives. Our evaluations show that this unified pretraining objective is better suited for a wider range of downstream tasks. Second, along the data axis, we pretrain on a curated dataset of 1 billion high-quality human images and improve the quality and quantity of task annotations. Third, architecturally, we incorporate advances from frontier models that enable longer training schedules with improved stability. Our 4K models adopt windowed attention to reason over longer spatial context and are pretrained with 2K output resolution. Sapiens2 sets a new state-of-the-art and improves over the first generation on pose (+4 mAP), body-part segmentation (+24.3 mIoU), normal estimation (45.6% lower angular error) and extends to new tasks such as pointmap and albedo estimation. Code: https://github.com/facebookresearch/sapiens2

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

Sapiens2 scales up the prior Sapiens work with bigger models, higher resolution, and a unified masked-plus-contrastive pretraining on a 1B human image set, but the gains are hard to attribute without ablations that fix data and scale.

read the letter

Sapiens2 scales the original Sapiens model into a family of 0.4B to 5B parameter transformers that run at native 1K resolution with 4K hierarchical variants. The main technical moves are a unified pretraining objective that combines masked image reconstruction and self-distilled contrastive learning, a curated 1B-image human dataset, and windowed attention to handle longer context at higher resolutions. It also adds tasks such as pointmap and albedo estimation while claiming concrete lifts on the old ones: +4 mAP pose, +24.3 mIoU segmentation, and 45.6% lower angular error on normals. Code is released, which helps anyone who wants to test the models directly. The paper does a straightforward job of extending the prior line of work to higher fidelity outputs and more tasks. The numbers look useful for people who need dense human predictions in graphics or robotics pipelines. The soft spot is exactly the one the stress-test flagged. The abstract states that the unified objective is better suited for a wider range of downstream tasks, yet it gives no ablations that keep data volume, curation quality, and model scale fixed. Without those controls it is difficult to know whether the reported improvements come from the new pretraining recipe or simply from training larger models on more carefully filtered human images. Evaluation protocols, baselines, and error bars are also not described in the abstract, so the SOTA claims cannot be checked at face value. This paper is for researchers who already work on human-centric vision and large-scale pretraining. Anyone building pose, segmentation, or normal estimators will find the benchmark numbers and the released code worth examining. It is not a foundational methods paper, but the scale and the task extensions make it worth a careful read. I would send it to peer review. Referees can ask for the missing ablations and verify the evaluation details; the core empirical claims are concrete enough to justify the time.

Referee Report

3 major / 2 minor

Summary. The paper introduces Sapiens2, a family of high-resolution (native 1K, hierarchical 4K) transformer models ranging from 0.4B to 5B parameters for human-centric vision. It claims that combining masked image reconstruction with self-distilled contrastive pretraining on a curated 1B-image human dataset, together with architectural advances (longer stable training, windowed attention), yields substantial gains over the prior Sapiens generation: +4 mAP on pose, +24.3 mIoU on body-part segmentation, 45.6% lower angular error on normals, plus new tasks (pointmap, albedo). The work positions the unified pretraining objective as better suited for dense prediction and zero/few-shot settings.

Significance. If the reported gains hold under controlled conditions, Sapiens2 would represent a meaningful advance in scalable, high-fidelity human-centric perception, particularly for applications needing both low-level detail and semantic robustness at high resolution. The extension to new dense tasks and the 4K variants are practically useful. However, the absence of isolating controls makes it difficult to attribute improvements to the proposed pretraining or architecture rather than dataset scale and curation.

major comments (3)

§4 (Experiments) and abstract: The SOTA claims and quantitative improvements (+4 mAP, +24.3 mIoU, 45.6% error reduction) are presented without any description of baselines, evaluation protocols, error bars, data splits, or ablation studies. This prevents verification of whether the unified pretraining objective, rather than the 1B-image dataset scale or undisclosed curation, drives the gains.
§3 (Pretraining): The central assertion that 'this unified pretraining objective is better suited for a wider range of downstream tasks' lacks any controlled comparison holding model scale, data volume, and curation fixed. Without such ablations, the attribution of improvements to masked reconstruction plus self-distilled contrastive learning versus simpler scaling remains untested.
§4.2 (Architectural changes and 4K models): The use of windowed attention for longer context and 2K pretraining resolution is described, but no ablation isolates its contribution to the reported metrics or stability gains relative to standard attention at equivalent compute.

minor comments (2)

The abstract and introduction would benefit from explicit pointers to the specific tables or figures that support each quantitative claim (e.g., the exact table reporting the +4 mAP pose result).
Model parameter counts and pretraining dataset details are mentioned but not tabulated with corresponding downstream performance; a summary table linking scale to each task would improve clarity.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the detailed and constructive review. We address each major comment below, clarifying the current manuscript content and outlining revisions to improve transparency and add supporting analyses where feasible.

read point-by-point responses

Referee: §4 (Experiments) and abstract: The SOTA claims and quantitative improvements (+4 mAP, +24.3 mIoU, 45.6% error reduction) are presented without any description of baselines, evaluation protocols, error bars, data splits, or ablation studies. This prevents verification of whether the unified pretraining objective, rather than the 1B-image dataset scale or undisclosed curation, drives the gains.

Authors: We agree that the experimental section would benefit from expanded descriptions of evaluation protocols, data splits, and error bars to aid reproducibility. In the revised manuscript we will add these details explicitly in §4, including the precise benchmark splits and reporting conventions used for each task. The reported gains are measured against the prior Sapiens model (same architecture family) as well as other published state-of-the-art methods on the respective public benchmarks; these comparisons are already tabulated but will be described more fully in text. Full-scale ablations that hold the 1 B-image curated dataset fixed while varying only the pretraining objective are computationally prohibitive, but we will include smaller-scale controlled experiments (e.g., 100 M images) to provide additional evidence on the contribution of the unified objective versus data scale. revision: partial
Referee: §3 (Pretraining): The central assertion that 'this unified pretraining objective is better suited for a wider range of downstream tasks' lacks any controlled comparison holding model scale, data volume, and curation fixed. Without such ablations, the attribution of improvements to masked reconstruction plus self-distilled contrastive learning versus simpler scaling remains untested.

Authors: We acknowledge that a controlled comparison at full scale would strengthen the claim. Repeating the entire 1 B-image pretraining run with alternative objectives is resource-intensive and was not performed. The manuscript motivates the unified objective by the complementary signals it provides (low-level detail from masked reconstruction and semantic robustness from self-distilled contrastive learning), and the downstream results across dense-prediction and zero/few-shot tasks are consistent with this motivation. In revision we will add a limitations paragraph discussing the absence of full-scale ablations and include reduced-scale experiments that hold model size and data volume fixed while varying the pretraining objective. revision: partial
Referee: §4.2 (Architectural changes and 4K models): The use of windowed attention for longer context and 2K pretraining resolution is described, but no ablation isolates its contribution to the reported metrics or stability gains relative to standard attention at equivalent compute.

Authors: We will add a dedicated ablation in the revised §4.2 that directly compares windowed attention against standard global attention under matched compute budgets. The study will report effects on training stability (loss curves and convergence speed) as well as downstream metrics for both the 1 K and 4 K model variants, thereby isolating the architectural contribution. revision: yes

Circularity Check

0 steps flagged

No circularity: purely empirical training and benchmarking

full rationale

The paper reports results from large-scale pretraining and fine-tuning of transformer models on a 1B-image dataset, with performance measured on standard external benchmarks (pose, segmentation, normals, etc.). No equations, fitted parameters, or derivations are presented that could reduce to inputs by construction. Improvements over Sapiens1 are stated as measured outcomes on held-out tasks rather than self-defined or self-cited necessities. Self-reference to prior work is present but does not carry the central claim; the work is self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

The central claims rest on standard transformer assumptions, the effectiveness of the chosen pretraining mix, and the quality of the curated 1B human image dataset; no new physical or mathematical axioms are introduced.

free parameters (2)

model parameter counts
Chosen range from 0.4B to 5B to explore scaling; specific values selected by authors.
pretraining dataset size
1 billion images curated by authors; exact filtering criteria not detailed in abstract.

axioms (1)

domain assumption Transformer architectures with windowed attention can stably train at 2K-4K resolutions for human images.
Invoked to justify the 4K hierarchical variant and longer training schedules.

pith-pipeline@v0.9.0 · 5560 in / 1363 out tokens · 35065 ms · 2026-05-09T21:56:36.355571+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

30 extracted references · 23 canonical work pages · 14 internal anchors

[1]

GQA: Training Generalized Multi-Query Transformer Models from Multi-Head Checkpoints

Joshua Ainslie, James Lee-Thorp, Michiel De Jong, Yury Zemlyanskiy, Federico Lebr´on, and Sumit Sanghai. Gqa: Training generalized multi-query transformer models from multi-head check- points.arXiv preprint arXiv:2305.13245,

work page internal anchor Pith review arXiv
[2]

Checkerboard artifact free sub-pixel convolution: A note on sub-pixel convolution, resize con- volution and convolution resize.arXiv preprint arXiv:1707.02937,

Andrew Aitken, Christian Ledig, Lucas Theis, Jose Caballero, Zehan Wang, and Wenzhe Shi. Checkerboard artifact free sub-pixel convolution: A note on sub-pixel convolution, resize con- volution and convolution resize.arXiv preprint arXiv:1707.02937,

work page arXiv
[3]

V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning

Mido Assran, Adrien Bardes, David Fan, Quentin Garrido, Russell Howes, Matthew Muckley, Am- mar Rizvi, Claire Roberts, Koustuv Sinha, Artem Zholus, et al. V-jepa 2: Self-supervised video models enable understanding, prediction and planning.arXiv preprint arXiv:2506.09985,

work page internal anchor Pith review arXiv
[4]

Loss functions in the era of semantic segmentation: A survey and outlook,

Reza Azad, Moein Heidary, Kadir Yilmaz, Michael H¨uttemann, Sanaz Karimijafarbigloo, Yuli Wu, Anke Schmeink, and Dorit Merhof. Loss functions in the era of semantic segmentation: A survey and outlook.arXiv preprint arXiv:2312.05391,

work page arXiv
[5]

BEiT: BERT Pre-Training of Image Transformers

Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254,

work page internal anchor Pith review arXiv
[6]

Revisiting Feature Prediction for Learning Visual Representations from Video

Adrien Bardes, Quentin Garrido, Jean Ponce, Xinlei Chen, Michael Rabbat, Yann LeCun, Mahmoud Assran, and Nicolas Ballas. Revisiting feature prediction for learning visual representations from video.arXiv preprint arXiv:2404.08471,

work page internal anchor Pith review arXiv
[7]

Depth Pro: Sharp Monocular Metric Depth in Less Than a Second

Aleksei Bochkovskii, Ama ˜AG ¸ l Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R Richter, and Vladlen Koltun. Depth pro: Sharp monocular metric depth in less than a second. arXiv preprint arXiv:2410.02073,

work page internal anchor Pith review arXiv
[8]

Perception Encoder: The best visual embeddings are not at the output of the network

Daniel Bolya, Po-Yao Huang, Peize Sun, Jang Hyun Cho, Andrea Madotto, Chen Wei, Tengyu Ma, Jiale Zhi, Jathushan Rajasegaran, Hanoona Rasheed, et al. Perception encoder: The best visual embeddings are not at the output of the network.arXiv preprint arXiv:2504.13181,

work page internal anchor Pith review arXiv
[9]

A simple framework for contrastive learning of visual representations

Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. InInternational conference on machine learning, pp. 1597–1607. PmLR, 2020a. Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E Hinton. Big self-supervised models are strong semi-supervised learn...

2026
[10]

Meta clip 2: A worldwide scaling recipe.arXiv preprint arXiv:2507.22062,

Yung-Sung Chuang, Yang Li, Dong Wang, Ching-Feng Yeh, Kehan Lyu, Ramya Raghavendra, James Glass, Lifei Huang, Jason Weston, Luke Zettlemoyer, et al. Meta clip 2: A worldwide scaling recipe.arXiv preprint arXiv:2507.22062,

work page arXiv
[11]

An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale

Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929,

work page internal anchor Pith review Pith/arXiv arXiv 2010
[12]

Convmae: Masked convolution meets masked autoencoders,

Peng Gao, Teli Ma, Hongsheng Li, Ziyi Lin, Jifeng Dai, and Yu Qiao. Convmae: Masked convolu- tion meets masked autoencoders.arXiv preprint arXiv:2205.03892,

work page arXiv
[13]

Query-Key Normalization for

Alex Henry, Prudhvi Raj Dachapally, Shubham Pawar, and Yuxuan Chen. Query-key normalization for transformers.arXiv preprint arXiv:2010.04245,

work page arXiv 2010
[14]

Scaling Laws for Neural Language Models

Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models.arXiv preprint arXiv:2001.08361,

work page internal anchor Pith review Pith/arXiv arXiv 2001
[15]

Sapiens: Foundation for human vision models

12 Published as a conference paper at ICLR 2026 Rawal Khirodkar, Timur Bagautdinov, Julieta Martinez, Su Zhaoen, Austin James, Peter Selednik, Stuart Anderson, and Shunsuke Saito. Sapiens: Foundation for human vision models. InEuro- pean Conference on Computer Vision, pp. 206–228. Springer,

2026
[16]

Geoman: Temporally consistent human geometry estimation using image-to-video diffusion.arXiv preprint arXiv:2505.23085, 2025a

Gwanghyun Kim, Xueting Li, Ye Yuan, Koki Nagano, Tianye Li, Jan Kautz, Se Young Chun, and Umar Iqbal. Geoman: Temporally consistent human geometry estimation using image-to-video diffusion.arXiv preprint arXiv:2505.23085, 2025a. Hoon Kim, Minje Jang, Wonjun Yoon, Jisoo Lee, Donghyun Na, and Sanghyun Woo. Switchlight: Co-design of physics-driven architectu...

work page arXiv
[17]

Decoupled Weight Decay Regularization

Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization.arXiv preprint arXiv:1711.05101,

work page internal anchor Pith review Pith/arXiv arXiv
[18]

Blending is all you need: Cheaper, better alternative to trillion-parameters llm.arXiv preprint arXiv:2401.02994,

Xiaoding Lu, Zongyi Liu, Adian Liusie, Vyas Raina, Vineet Mudupalli, Yuwen Zhang, and William Beauchamp. Blending is all you need: Cheaper, better alternative to trillion-parameters llm.arXiv preprint arXiv:2401.02994,

work page arXiv
[19]

The llama 4 herd: The beginning of a new era of natively multimodal ai innovation

AI Meta. The llama 4 herd: The beginning of a new era of natively multimodal ai innovation. https://ai. meta. com/blog/llama-4-multimodal-intelligence/, checked on, 4(7):2025,

2025
[20]

DINOv2: Learning Robust Visual Features without Supervision

Maxime Oquab, Timoth ´ee Darcet, Th ´eo Moutakanni, Huy V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision.arXiv preprint arXiv:2304.07193,

work page internal anchor Pith review Pith/arXiv arXiv
[21]

High-fidelity facial albedo estimation via texture quantization.arXiv preprint arXiv:2406.13149,

Zimin Ran, Xingyu Ren, Xiang An, Kaicheng Yang, Xiangzi Dai, Ziyong Feng, Jia Guo, Linchao Zhu, and Jiankang Deng. High-fidelity facial albedo estimation via texture quantization.arXiv preprint arXiv:2406.13149,

work page arXiv
[22]

Hiera: A hierarchical vision transformer without the bells-and-whistles

13 Published as a conference paper at ICLR 2026 Chaitanya Ryali, Yuan-Ting Hu, Daniel Bolya, Chen Wei, Haoqi Fan, Po-Yao Huang, Vaibhav Ag- garwal, Arkabandhu Chowdhury, Omid Poursaeed, Judy Hoffman, et al. Hiera: A hierarchical vision transformer without the bells-and-whistles. InInternational conference on machine learn- ing, pp. 29441–29454. PMLR,

2026
[23]

David: Data-efficient and accurate vision models from syn- thetic data.arXiv preprint arXiv:2507.15365,

Fatemeh Saleh, Sadegh Aliakbarian, Charlie Hewitt, Lohit Petikam, Antonio Criminisi, Thomas J Cashman, Tadas Baltruˇsaitis, et al. David: Data-efficient and accurate vision models from syn- thetic data.arXiv preprint arXiv:2507.15365,

work page arXiv
[24]

GLU Variants Improve Transformer

Noam Shazeer. Glu variants improve transformer.arXiv preprint arXiv:2002.05202,

work page internal anchor Pith review arXiv 2002
[25]

DINOv3

Oriane Sim´eoni, Huy V V o, Maximilian Seitzer, Federico Baldassarre, Maxime Oquab, Cijo Jose, Vasil Khalidov, Marc Szafraniec, Seungeun Yi, Micha ¨el Ramamonjisoa, et al. Dinov3.arXiv preprint arXiv:2508.10104,

work page internal anchor Pith review Pith/arXiv arXiv
[26]

LLaMA: Open and Efficient Foundation Language Models

Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timoth ´ee Lacroix, Baptiste Rozi `ere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models.arXiv preprint arXiv:2302.13971,

work page internal anchor Pith review Pith/arXiv arXiv
[27]

Deep learning-based human pose estimation: A survey.ACM computing surveys, 56(1):1–37,

14 Published as a conference paper at ICLR 2026 Ce Zheng, Wenhan Wu, Chen Chen, Taojiannan Yang, Sijie Zhu, Ju Shen, Nasser Kehtarnavaz, and Mubarak Shah. Deep learning-based human pose estimation: A survey.ACM computing surveys, 56(1):1–37,

2026
[28]

iBOT: Image BERT Pre-Training with Online Tokenizer

Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer.arXiv preprint arXiv:2111.07832,

work page internal anchor Pith review arXiv
[29]

For instance, we pretrain the SAPIENS2–1B (embed dim1536,40layers,24 heads, patch size16, final norm with [CLS]) at1024×768

15 Published as a conference paper at ICLR 2026 A APPENDIX A.1 PRETRAINING A.1.1 IMPLEMENTATIONDETAILS We use the dense-probing evaluations as the final metrics to guide any design decisions during the pretraining stage. For instance, we pretrain the SAPIENS2–1B (embed dim1536,40layers,24 heads, patch size16, final norm with [CLS]) at1024×768. Training us...

2026
[30]

Blockwise Masking Patchwise Masking Figure 11: We randomly mix blockwise and patchwise masking to provide coarse occlusions

At1024×768(64×48=3072 patches), this masks∼2304patches per image, yielding coarse occlusions that regularize MAE while leaving sufficient context for contrastive learning. Blockwise Masking Patchwise Masking Figure 11: We randomly mix blockwise and patchwise masking to provide coarse occlusions. For MAE pre- training at high resolution (1024), we use a75%...

2026