On the Design Fundamentals of Pixel Text Representation Learning
Pixel Linguist II is presented as a stronger native-resolution encoder for reading text directly from images.
The paper argues that robust pixel-text learning depends on variable resolutions and font sizes, natural image-text grounding, layout-aware rendering, and a two-stage multilingual curriculum. The authors trained the model on 280M examples with on-the-fly rendering and unified contrastive grounding. They report state-of-the-art results across English, cross-lingual, and multilingual Visual STS and ViDoRe benchmarks, plus stronger downstream MLLM evaluation. The model also remains robust with 80% visual token compression, which the authors frame as useful for optical context compression. ArXiv · AI/CL/LG's note
The paper argues that robust pixel-text learning depends on variable resolutions and font sizes, natural image-text grounding, layout-aware rendering, and a two-stage multilingual curriculum. The authors trained the model on 280M examples with on-the-fly rendering and unified contrastive grounding. They report state-of-the-art results across English, cross-lingual, and multilingual Visual STS and ViDoRe benchmarks, plus stronger downstream MLLM evaluation. The model also remains robust with 80% visual token compression, which the authors frame as useful for optical context compression. ArXiv · AI/CL/LG's note
score 4