SigLIP 2: Multilingual Vision-Language Encoders with Improved Semantic Understanding, Localization, and Dense Features

By Michael Tschannen, Alexey Gritsenko, Xiao Wang, M. Naeem, Ibrahim M. Alabdulmohsin, Nikhil Parthasarathy, Talfan Evans, Lucas Beyer, Ye Xia, Basil Mustafa and 4 more. arXiv.org, 2025

1,063 citations recorded by the source corpus.

We introduce SigLIP 2, a family of new multilingual vision-language encoders that build on the success of the original SigLIP. In this second iteration, we extend the original image-text training objective with several prior, independently developed techniques into a unified recipe -- this includes captioning-based pretraining, self-supervised losses (self-distillation, masked prediction) and online data curation. With these changes, SigLIP 2 models outperform their SigLIP counterparts at all model scales in core capabilities, including zero-shot classification, image-text retrieval, and transfer performance when extracting visual representations for Vision-Language Models (VLMs). Furthermore, the new training recipe leads to significant improvements on localization and dense prediction tasks. We also train variants which support multiple resolutions and preserve the input's native aspect ratio. Finally, we train on a more diverse data-mixture that includes de-biasing techniques, leading to much better multilingual understanding and improved fairness. To allow users to trade off inference cost with performance, we release model checkpoints at four sizes: ViT-B (86M), L (303M), So400m (400M), and g (1B).

Subfields: Computer Vision and Pattern Recognition

View Paper: https://www.semanticscholar.org/paper/5bbb904e85ced452f44d9ed9559297334298c407

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Vision and Pattern Recognition

1. Rongyao Fang
2. Jingren Zhou
3. Shuai Bai
4. Lingchen Meng
5. Rui Men
6. Qidong Huang
7. Zesen Cheng

Citations in the corpus

• Qwen3-Omni Technical Report (2025, positive): “Furthermore, we employ the vision encoder from Qwen3-VL, initialized from SigLIP2-So400m (Tschannen et al., 2025) with approximately 543 million parameters, enabling handling of both image and video inputs.”
• V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning (2025, neutral): “We observe that V-JEPA 2 exhibits competitive performance in the frozen setup, outperforming DINOv2, SigLIP, and Perception Encoder (PE) in all of the tested benchmarks (Table 6) except PerceptionTest where V-JEPA 2 slightly underperforms SigLIP and PE.”
• Emerging Properties in Unified Multimodal Pretraining (2025, positive): “We adopt SigLIP2-so400m/14 [75] with a fixed 384 -resolution as the initialization of the ViT encoder.”
• Qwen-Image Technical Report (2025, neutral): “Janus-Pro (Chen et al., 2025b) 0.019 0.006 BLIP3-o (Chen et al., 2025a) 0.021 0.018 Kolors 2.0 (team, 2025) 0.258 0.329 BAGEL (Deng et al., 2025) 0.373 0.310 OmniGen2 (Wu et al., 2025b) 0.561 0.059 X-Omni (Geng et al., 2025) 0.900 0.814 HiDream-I1-Full (Cai et al., 2025) 0.543 0.024 FLUX.1 [Dev] (Bl”
• Qwen3-VL Technical Report (2025, positive): “Vision Encoder : We utilize the SigLIP-2 architecture (Tschannen et al., 2025) as our vision encoder and continue training it with dynamic input resolutions, initialized from official pretrained checkpoints.”
• Eigenfunction Extraction for Ordered Representation Learning (2025, unknown): “These methods can also be applied to multi-modal learning, where shared representation spaces are learned from information from aligned data from different modalities (Radford et al., 2021; Jia et al., 2021; Yuan et al., 2021; Li et al., 2022; Chuang et al., 2025; Tschannen et al., 2025).”
• CLeaR: A Unified Framework for Resolving the Leakage-Degradation Dilemma in Style Transfer (2026, unknown): “We add SigLIP [39] and Nomic [19] as additional VFMs.”
• VaRS-Doc: Interpretation-Aware Variant Representations via Latent Self-Probing for Visual Document Retrieval (2026, unknown): “Multimodal large language models have made rapid progress in understanding visually rich documents (Fu et al. 2025; Alayrac et al. 2022; Li et al. 2022; Tschannen et al. 2025).”
• LightSTAR: Efficient Visual Document Retrieval via Lightweight Selection with Vision-Adaptive Refinement (2026, unknown): “Vision-language models have achieved remarkable progress in document understanding tasks [10,12–20,29,41,45–47].”
• jina-embeddings-v5-omni: Geometry-preserving Embeddings via Locked Aligned Towers (2026, unknown): “CLIP [27] established contrastive image–text embedding with separatelyencodedimage andtexttowers,and SigLIP[38],SigLIP2[32], and EVA-CLIP [10] refine this paradigm through improved losses, data, and visual training recipes.”
• Visual Product Search Benchmark (2026, unknown): “The suffix SO denotes a shape-optimized variant, which is trained to be more robust to variations in object scale, aspect ratio, and image composition [33].”
• Nemotron ColEmbed V2: Top-Performing Late Interaction embedding models for Visual Document Retrieval (2026, unknown): “While highly efficient and common in early text-only retrieval [35] or CLIP/SigLIP-style VLMs [14, 15], the single-vector representation bottleneck is often limited to capture fine-grained lexical, visual, or layout-dependent cues.”
• Language-Agnostic Visual Embeddings for Cross-Script Handwriting Retrieval (2026, unknown): “Conversely, embedding-based methods map images directly into a latent space to bypass explicit transcription [Yang et al. , 2022; Zhou et al. , 2024; Tschannen et al. , 2025; Zhang et al. , 2025], significantly reducing structural complexity.”
• AURA: A Fine-Grained Benchmark and Decomposed Metric for Audio-Visual Reasoning (2025, unknown): “…and cross-modal integration [Li et al. 2021, Ghosh et al. 2024, Nag, Goswami, and Karanam 2024, Zhang et al. 2022, Zhai et al. 2023, Mittal et al. 2024, Tschannen et al. 2025, Pra-manick et al. 2023b, Guo et al. 2025, Deb et al. 2023, Gong et al. 2025, Pramanick et al. 2023c, Sun et al. 2025a].”