Practical overview
When it fits
Use Seg Zero when you want research code for reasoning-guided image segmentation, or a starting point for experiments with vision-language models, segmentation rewards and reinforcement-learning training.
Who it is for
- Researchers studying multimodal reasoning and visual segmentation
- Developers experimenting with reinforcement-learning training for vision-language models
- Teams evaluating prompted segmentation on ReasonSeg data
Documented uses
- Generate a segmentation mask from an image and text prompt while viewing the model's reasoning trace.
- Evaluate a model on ReasonSeg data and inspect gIoU output.
- Explore reinforcement-learning training methods for visual perception and reasoning.
Requirements & limits
The repository is public and licensed under Apache-2.0. Its installation instructions specify Linux-oriented Python tooling, including Conda, PyTorch and torchvision.
- The inference instructions download VisionReasoner-7B, while the repository also lists Seg-Zero-7B as a separate model.
- The README describes multi-object segmentation in the current version and provides a command to revert to an earlier single-object version.
- The suggested hardware for 7B training is four 80 GB GPUs or eight 46 GB GPUs.
- Benchmark results for Seg-Zero use different checkpoints across benchmarks, which affects direct comparisons.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Run inference on an image and text prompt
Set up the repository, download the VisionReasoner-7B checkpoint and run the documented multi-object inference script.
Before you start
- Python 3.12 with Conda
- PyTorch 2.6.0 and torchvision 0.21.0
- Git LFS
- The VisionReasoner-7B model files
- A local image and text prompt for custom inference
Steps
Clone the Seg-Zero repository and create and activate a Conda environment using Python 3.12.
Install the documented PyTorch and torchvision versions, then install the repository in editable mode.
Create a pretrained_models directory, initialize Git LFS and clone the VisionReasoner-7B model repository into it.
Run the multi-object inference script; provide --image_path and --text to use a custom image and prompt.
Expected outcome
The script prints a reasoning process in the command line and presents a segmentation mask in the inference_scripts folder.
Guide limitations
- The README gives a suggested model-download path but does not specify hardware requirements for inference.
- For a 7B training setup, the repository recommends four 80 GB GPUs or eight 46 GB GPUs; this is a training recommendation, not an inference requirement.
Evaluate with the ReasonSeg script
Run the repository's documented ReasonSeg evaluation command and inspect its gIoU output.
Before you start
- A Seg-Zero repository environment
- The evaluation data identified as ReasonSeg-Test and ReasonSeg-Val
- A GPU setup appropriate for the chosen batch size
Steps
Use the documented ReasonSeg-Test and ReasonSeg-Val evaluation data.
Run the evaluation script with bash evaluation_scripts/eval_reasonseg_visionreasoner.sh.
Adjust --batch_size in the shell script to suit the available GPU.
Expected outcome
The evaluation script reports gIoU in the command line.
Guide limitations
- Seg-Zero's reported best results across benchmarks use different checkpoints; the README recommends evaluating benchmarks with one checkpoint when making comparisons.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 15 October 2026
- Original website title
- JIA-Lab-research/Seg-Zero · Official homepage title · retained in its source language
Related options
VisionReasoner — A related project from the same repository organization; Seg Zero's repository provides its official training code and recommends it for evaluation across more tasks and benchmarks
Explore project →