---
license: mit
pipeline_tag: image-segmentation
library_name: transformers
---

# MLLMSeg: Unlocking the Potential of MLLMs in Referring Expression Segmentation via a Light-weight Mask Decoder

This repository contains the `MLLMSeg_InternVL2_5_4B_RES` model, which was presented in the paper [Unlocking the Potential of MLLMs in Referring Expression Segmentation via a Light-weight Mask Decoder](https://huggingface.co/papers/2508.04107).

**MLLMSeg** is a novel framework designed for Referring Expression Segmentation (RES). It fully exploits the inherent visual detail features encoded in Multimodal Large Language Model (MLLM) vision encoders without introducing an extra visual encoder. The model employs a detail-enhanced and semantic-consistent feature fusion module (DSFF) and a light-weight mask decoder (with only 34M network parameters) to achieve precise mask prediction. This approach strikes a better balance between performance and cost compared to existing methods.

Code: https://github.com/jcwang0602/MLLMSeg

## Quick Start (Inference)

### Installation

First, install the `transformers` library and other dependencies. For a complete installation guide, please refer to the [official GitHub repository](https://github.com/jcwang0602/MLLMSeg) and the [InternVL2 documentation](https://internvl.readthedocs.io/en/latest/get_started/installation.html).

```bash
conda create -n mllmseg python==3.10.18 -y
conda activate mllmseg
pip install torch==2.5.1 torchvision==0.20.1 --index-url https://download.pytorch.org/whl/cu118
pip install -r requirements.txt
pip install flash-attn==2.3.6 --no-build-isolation # Note: need gpu to install
```

### Inference Example

Here's an example to perform inference with the model:

```python
import numpy as np
import torch
import torchvision.transforms as T
from PIL import Image
from torchvision.transforms.functional import InterpolationMode
from transformers import AutoModel, AutoTokenizer

IMAGENET_MEAN = (0.485, 0.456, 0.406)
IMAGENET_STD = (0.229, 0.224, 0.225)

def build_transform(input_size):
    MEAN, STD = IMAGENET_MEAN, IMAGENET_STD
    transform = T.Compose([
        T.Lambda(lambda img: img.convert('RGB') if img.mode != 'RGB' else img),
        T.Resize((input_size, input_size), interpolation=InterpolationMode.BICUBIC),
        T.ToTensor(),
        T.Normalize(mean=MEAN, std=STD)
    ])
    return transform

def find_closest_aspect_ratio(aspect_ratio, target_ratios, width, height, image_size):
    best_ratio_diff = float('inf')
    best_ratio = (1, 1)
    area = width * height
    for ratio in target_ratios:
        target_aspect_ratio = ratio[0] / ratio[1]
        ratio_diff = abs(aspect_ratio - target_aspect_ratio)
        if ratio_diff < best_ratio_diff:
            best_ratio_diff = ratio_diff
            best_ratio = ratio
        elif ratio_diff == best_ratio_diff:
            if area > 0.5 * image_size * image_size * ratio[0] * ratio[1]:
                best_ratio = ratio
    return best_ratio

def dynamic_preprocess(image, min_num=1, max_num=12, image_size=448, use_thumbnail=False):
    orig_width, orig_height = image.size
    aspect_ratio = orig_width / orig_height

    # calculate the existing image aspect ratio
    target_ratios = set(
        (i, j) for n in range(min_num, max_num + 1) for i in range(1, n + 1) for j in range(1, n + 1) if
        i * j <= max_num and i * j >= min_num)
    target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1])

    # find the closest aspect ratio to the target
    target_aspect_ratio = find_closest_aspect_ratio(
        aspect_ratio, target_ratios, orig_width, orig_height, image_size)

    # calculate the target width and height
    target_width = image_size * target_aspect_ratio[0]
    target_height = image_size * target_aspect_ratio[1]
    blocks = target_aspect_ratio[0] * target_aspect_ratio[1]

    # resize the image
    resized_img = image.resize((target_width, target_height))
    processed_images = []
    for i in range(blocks):
        box = (
            (i % (target_width // image_size)) * image_size,
            (i // (target_width // image_size)) * image_size,
            ((i % (target_width // image_size)) + 1) * image_size,
            ((i // (target_width // image_size)) + 1) * image_size
        )
        # split the image
        split_img = resized_img.crop(box)
        processed_images.append(split_img)
    assert len(processed_images) == blocks
    if use_thumbnail and len(processed_images) != 1:
        thumbnail_img = image.resize((image_size, image_size))
        processed_images.append(thumbnail_img)
    return processed_images

def load_image(image_file, input_size=448, max_num=12):
    image = Image.open(image_file).convert('RGB')
    transform = build_transform(input_size=input_size)
    images = dynamic_preprocess(image, image_size=input_size, use_thumbnail=True, max_num=max_num)
    pixel_values = [transform(image) for image in images]
    pixel_values = torch.stack(pixel_values)
    return pixel_values

# Load the model and tokenizer
path = 'jcwang0602/MLLMSeg_InternVL2_5_4B_RES'
model = AutoModel.from_pretrained(
    path,
    torch_dtype=torch.bfloat16,
    low_cpu_mem_usage=True,
    trust_remote_code=True).eval().cuda()
tokenizer = AutoTokenizer.from_pretrained(path, trust_remote_code=True, use_fast=False)

# Prepare image and question (replace './examples/images/web_dfacd48d-d2c2-492f-b94c-41e6a34ea99f.png' with your image path)
pixel_values = load_image('./examples/images/web_dfacd48d-d2c2-492f-b94c-41e6a34ea99f.png', max_num=6).to(torch.bfloat16).cuda()
generation_config = dict(max_new_tokens=1024, do_sample=True)

question = "In the screenshot of this web page, please give me the coordinates of the element I want to click on according to my instructions(with point).\
\\\"'Champions League' link\\\""

# Chat with the model
response, history = model.chat(tokenizer, pixel_values, question, generation_config, history=None, return_history=True)
print(f'User: {question}\
Assistant: {response}')
```

## Performance Metrics

The following tables showcase the performance of MLLMSeg on various benchmarks, as presented in the original repository:

### Referring Expression Segmentation
<img src="https://github.com/jcwang0602/MLLMSeg/raw/main/assets/tab_res.png" width="800">

### Referring Expression Comprehension
<img src="https://jcwang0602/MLLMSeg/raw/main/assets/tab_rec.png" width="800">

### Generalized Referring Expression Segmentation
<img src="https://github.com/jcwang0602/MLLMSeg/raw/main/assets/tab_gres.png" width="800">

## Visualization

Visual examples of MLLMSeg's performance:

### Referring Expression Segmentation
<img src="https://github.com/jcwang0602/MLLMSeg/raw/main/assets/res.png" width="800">

### Referring Expression Comprehension
<img src="https://github.com/jcwang0602/MLLMSeg/raw/main/assets/rec.png" width="800">

### Generalized Referring Expression Segmentation
<img src="https://github.com/jcwang0602/MLLMSeg/raw/main/assets/gres.png" width="800">

## Citation

If our work is useful for your research, please consider citing:
```bibtex
@misc{wang2025unlockingpotentialmllmsreferring,
      title={Unlocking the Potential of MLLMs in Referring Expression Segmentation via a Light-weight Mask Decoder}, 
      author={Jingchao Wang and Zhijian Wu and Dingjiang Huang and Yefeng Zheng and Hong Wang},
      year={2025},
      eprint={2508.04107},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2508.04107}, 
}
```