Sa2VA is an MLLM capable of question answering, visual prompt understanding, and dense object segmentation at both image and video levels. It achieves comparable performance to SOTA MLLMs Qwen2.5-VL and InternVL3 on question-answering benchmarks. Additionally, Sa2VA possesses the visual prompt understanding and dense object segmentation capabilities that SOTA MLLMs Qwen2.5-VL and InternVL3 lack. Sa2VA achieves SOTA performance on both image and video grounding and segmentation benchmarks.

Sa2VA Family

We built the Sa2VA series based on Qwen2.5/3-VL and InternVL2.5/3. In the following table, we provide some Sa2VA models built on Qwen2.5/3-VL and InternVL3.

Model Name	Base MLLM	Language Part	HF Link
Sa2VA-InternVL3-2B	InternVL3-2B	Qwen2.5-1.5B	🤗 link
Sa2VA-InternVL3-8B	InternVL3-8B	Qwen2.5-7B	🤗 link
Sa2VA-InternVL3-14B	InternVL3-14B	Qwen2.5-14B	🤗 link
Sa2VA-Qwen2_5-VL-3B	Qwen2.5-VL-3B-Instruct	Qwen2.5-3B	🤗 link
Sa2VA-Qwen2_5-VL-7B	Qwen2.5-VL-7B-Instruct	Qwen2.5-7B	🤗 link
Sa2VA-Qwen3-VL-2B	Qwen3-VL-2B-Instruct	Qwen3-1.7B	🤗 link
Sa2VA-Qwen3-VL-4B	Qwen3-VL-4B-Instruct	Qwen3-4B	🤗 link

Sa2VA Performance

Model Name	MME	MMBench	RefCOCO	RefCOCO+	RefCOCOg	MeVIS (val_u)	DAVIS
Sa2VA-InternVL3-2B	1631/559	79.8	81.4	75.7	80.3	53.9	74.5
Sa2VA-InternVL3-8B	1743/633	83.0	83.3	78.9	81.8	56.4	76.3
Sa2VA-InternVL3-14B	1746/724	84.3	83.6	79.9	83.6	59.2	76.6
Sa2VA-Qwen2_5-VL-3B	1533/572	78.4	79.6	74.0	77.1	51.6	73.4
Sa2VA-Qwen2_5-VL-7B	1552/676	84.5	82.4	77.5	81.5	56.4	79.4
Sa2VA-Qwen3-VL-2B	1541/520	79.0	80.2	75.1	78.5	53.9	74.8
Sa2VA-Qwen3-VL-4B	1660/655	86.3	81.7	77.4	80.0	57.1	75.9

Quick Start

We provide an example code to run Sa2VA using transformers.

python

import torch
from transformers import AutoProcessor, AutoModel
from PIL import Image
import numpy as np
import os

# load the model and processor
path = "ByteDance/Sa2VA-Qwen3-VL-4B"
model = AutoModel.from_pretrained(
    path,
    torch_dtype=torch.bfloat16,
    low_cpu_mem_usage=True,
    use_flash_attn=True,
    trust_remote_code=True).eval().cuda()
processor = AutoProcessor.from_pretrained(path, trust_remote_code=True, use_fast=False)

# for image chat
image_path = "/PATH/TO/IMAGE"
text_prompts = "<image>Please describe the image."
image = Image.open(image_path).convert('RGB')
input_dict = {
    'image': image,
    'text': text_prompts,
    'past_text': '',
    'mask_prompts': None,
    'processor': processor,
    }
return_dict = model.predict_forward(**input_dict)
answer = return_dict["prediction"] # the text format answer

# for image chat with segmentation output
image_path = "/PATH/TO/IMAGE"
text_prompts = "<image>Could you please give me a brief description of the image? Please respond with interleaved segmentation masks for the corresponding parts of the answer."
image = Image.open(image_path).convert('RGB')
input_dict = {
    'image': image,
    'text': text_prompts,
    'past_text': '',
    'mask_prompts': None,
    'processor': processor,
    }
return_dict = model.predict_forward(**input_dict)
answer = return_dict["prediction"] # the text format answer
masks = return_dict['prediction_masks']  # segmentation masks, list(np.array(1, h, w), ...)
    
# for chat with visual prompt (mask format) input
mask_prompts = np.load('/PATH/TO/pred_masks.npy') # np.array(n_prompts, h, w)
image_path = "/PATH/TO/IMAGE"
text_prompts = "<image>Can you provide me with a detailed description of the region in the picture marked by region1."
image = Image.open(image_path).convert('RGB')
input_dict = {
    'image': image,
    'text': text_prompts,
    'past_text': '',
    'mask_prompts': mask_prompts,
    'processor': processor,
    }
return_dict = model.predict_forward(**input_dict)
answer = return_dict["prediction"] # the text format answer

# for video chat
video_folder = "/PATH/TO/VIDEO_FOLDER"
images_paths = os.listdir(video_folder)
images_paths = [os.path.join(video_folder, image_path) for image_name in images_paths]
if len(images_paths) > 5:  # uniformly sample 5 frames
    step = (len(images_paths) - 1) // (5 - 1)
    images_paths = [images_paths[0]] + images_paths[1:-1][::step][1:] + [images_paths[-1]]
text_prompts = "<image>Please describe the video."
input_dict = {
    'video': images_paths,
    'text': text_prompts,
    'past_text': '',
    'mask_prompts': None,
    'processor': processor,
}
return_dict = model.predict_forward(**input_dict)
answer = return_dict["prediction"] # the text format answer


# for video chat with segmentation mask output
video_folder = "/PATH/TO/VIDEO_FOLDER"
images_paths = os.listdir(video_folder)
images_paths = [os.path.join(video_folder, image_path) for image_name in images_paths]
text_prompts = "<image>Please segment the person."
input_dict = {
    'video': images_paths,
    'text': text_prompts,
    'past_text': '',
    'mask_prompts': None,
    'processor': processor,
}
return_dict = model.predict_forward(**input_dict)
answer = return_dict["prediction"] # the text format answer
masks = return_dict['prediction_masks']  # segmentation masks, list(np.array(n_frames, h, w), ...)

Citation

If you find this project useful in your research, please consider citing:

BibTeX

@article{sa2va,
  title={Sa2VA: Marrying SAM2 with LLaVA for Dense Grounded Understanding of Images and Videos},
  author={Yuan, Haobo and Li, Xiangtai and Zhang, Tao and Huang, Zilong Huang and Xu, Shilin and Ji, Shunping and Tong, Yunhai and Qi, Lu and Feng, Jiashi and Yang, Ming-Hsuan},
  journal={arXiv preprint},
  year={2025}
}

ByteDance/Sa2VA-Qwen3-VL-2B

Sa2VA: Marrying SAM2 with LLaVA for Dense Grounded Understanding of Images and Videos

Introduction

Sa2VA Family

Sa2VA Performance

Quick Start

Citation

ByteDance/
Sa2VA-Qwen3-VL-2B