AI Trends

Trending story

Ant Group open-sources ARGenSeg inference and training code and checkpoints

1 reports1 story updatesUpdated Nov 17 12:41

Understand this story

AI overview

Ant Group's inclusionAI has open-sourced the inference code, training code, and the ARGenSeg-8B checkpoint for the NeurIPS 2025 paper ARGenSeg. The framework has the MLLM output visual tokens, which are then decoded into dense masks via a general-purpose VQ-VAE, eliminating the need for a dedicated segmentation head, and it uses next-scale prediction to generate visual tokens in parallel to reduce inference latency. On multiple segmentation datasets, its performance exceeds the previous SOTA, with significantly faster inference.

AI-generated from public reports

Latest development · Report

Ant inclusionAI Open-Sources ARGenSeg: Using an Autoregressive Image Generation Model for Image Segmentation

Developments

1 confirmed developments

  1. Ant Group open-sources ARGenSeg inference, training code, and checkpoints

    Source:English text is being prepared.

Reports and discussion timeline

Browse reports and discussion by time. Their evidence levels remain distinct.

Nov 17

  1. Ant inclusionAI: New GitHub RepositoryFirst-party

    Ant inclusionAI Open-Sources ARGenSeg: Using an Autoregressive Image Generation Model for Image Segmentation

    Ant inclusionAI has open-sourced the inference code, training code, and ARGenSeg-8B checkpoint for the NeurIPS 2025 paper ARGenSeg. The framework has an MLLM output visual tokens, which are then decoded into dense masks via a general-purpose VQ-VAE, eliminating the need for a dedicated segmentation head, and it employs next-scale prediction to generate visual tokens in parallel to reduce inference latency. On multiple segmentation datasets, its performance surpasses the previous SOTA, with significantly improved inference speed.

Related stories

  1. Related storiesinclusionAI open-sources Ming-UniVision and MingTok-Vision
  2. Related storiesinclusionAI open-sources the ZwZ model family and the ZoomBench benchmark
  3. Related storiesMiniMax releases the H3-Image image model and opens the weights
  4. Related storiesApple research team submits paper to arXiv disclosing the HALO system