# Tested model inference environment for RefCaptioner (Python 3.10).
torch==2.9.0
torchvision==0.24.0
transformers==4.57.6
accelerate==1.14.0
qwen-vl-utils==0.0.14
av==17.1.0
pillow==12.2.0
numpy==2.2.6
peft==0.19.1
safetensors==0.8.0
einops==0.8.2