--- license: apache-2.0 pipeline_tag: image-text-to-text library_name: transformers base_model: - OpenGVLab/InternVL3-8B base_model_relation: merge language: - multilingual tags: - Sa2VA - custom_code --- # Sa2VA: Marrying SAM2 with LLaVA for Dense Grounded Understanding of Images and Videos [\[📂 GitHub\]](https://github.com/bytedance/Sa2VA) [\[📜 Sa2VA paper\]](https://arxiv.org/abs/2501.04001) [\[🚀 Quick Start\]](#quick-start) ## Introduction 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](https://huggingface.co/OpenGVLab/InternVL3-2B) | [Qwen2.5-1.5B](https://huggingface.co/Qwen/Qwen2.5-1.5B) | [🤗 link](https://huggingface.co/ByteDance/Sa2VA-InternVL3-2B) | | Sa2VA-InternVL3-8B | [InternVL3-8B](https://huggingface.co/OpenGVLab/InternVL3-8B) | [Qwen2.5-7B](https://huggingface.co/Qwen/Qwen2.5-7B) | [🤗 link](https://huggingface.co/ByteDance/Sa2VA-InternVL3-8B) | | Sa2VA-InternVL3-14B | [InternVL3-14B](https://huggingface.co/OpenGVLab/InternVL3-14B) | [Qwen2.5-14B](https://huggingface.co/Qwen/Qwen2.5-14B) | [🤗 link](https://huggingface.co/ByteDance/Sa2VA-InternVL3-14B) | | Sa2VA-Qwen2_5-VL-3B | [Qwen2.5-VL-3B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-3B-Instruct) | [Qwen2.5-3B](https://huggingface.co/Qwen/Qwen2.5-3B) | [🤗 link](https://huggingface.co/ByteDance/Sa2VA-Qwen2_5-VL-3B) | | Sa2VA-Qwen2_5-VL-7B | [Qwen2.5-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct) | [Qwen2.5-7B](https://huggingface.co/Qwen/Qwen2.5-7B) | [🤗 link](https://huggingface.co/ByteDance/Sa2VA-Qwen2_5-VL-7B) | | Sa2VA-Qwen3-VL-4B | [Qwen3-VL-4B-Instruct](https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct) | [Qwen3-4B](https://huggingface.co/Qwen/Qwen3-4B) | [🤗 link](https://huggingface.co/ByteDance/Sa2VA-Qwen3-VL-4B) | ## 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-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 = "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 = "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 = "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 = "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 = "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} } ```