--- base_model: - Qwen/Qwen2.5-VL-7B-Instruct language: - en library_name: transformers license: apache-2.0 pipeline_tag: image-text-to-text tags: - gui - agent - gui-grounding - reinforcement-learning --- # InfiGUI-G1-7B This repository contains the InfiGUI-G1-7B model from the paper **[InfiGUI-G1: Advancing GUI Grounding with Adaptive Exploration Policy Optimization](https://arxiv.org/abs/2508.05731)**.

## Model Description The model is based on `Qwen2.5-VL-7B-Instruct` and is fine-tuned using our proposed **Adaptive Exploration Policy Optimization (AEPO)** framework. AEPO is a novel reinforcement learning method designed to enhance the model's **semantic alignment** for GUI grounding tasks. It overcomes the exploration bottlenecks of standard RLVR methods by integrating a multi-answer generation strategy with a theoretically-grounded adaptive reward function, enabling more effective and efficient learning for complex GUI interactions. ## Paper Overview A fundamental challenge for GUI agents is robustly grounding natural language instructions, which requires not only precise **spatial alignment** (locating elements accurately) but also correct **semantic alignment** (identifying the functionally appropriate element). While existing Reinforcement Learning with Verifiable Rewards (RLVR) methods have enhanced spatial precision, they often suffer from inefficient exploration. This "confidence trap" bottlenecks semantic alignment, preventing models from discovering correct actions for difficult semantic associations. To address this critical exploration problem, we introduce **InfiGUI-G1**, a series of models trained with **Adaptive Exploration Policy Optimization (AEPO)**. AEPO overcomes the exploration bottleneck by integrating a **multi-answer generation** strategy to explore a diverse set of candidate actions in a single forward pass. This exploration is guided by a theoretically-grounded **Adaptive Exploration Reward (AER)** function, derived from first principles of efficiency (η=U/C), which provides rich, informative learning signals to dynamically balance exploration and exploitation. ## Quick Start ### Installation First, install the required dependencies: ```bash pip install transformers qwen-vl-utils ```` ### Example ```python import json import math import torch import requests from io import BytesIO from PIL import Image, ImageDraw, ImageFont from transformers import Qwen2_5_VLForConditionalGeneration, AutoProcessor from qwen_vl_utils import process_vision_info, smart_resize MAX_IMAGE_PIXELS = 5600 * 28 * 28 def resize_image(width: int, height: int, max_pixels: int) -> tuple[int, int]: """ Resize image to fit within max_pixels constraint while maintaining aspect ratio. Applies smart_resize for final dimension optimization. """ current_pixels = width * height if current_pixels <= max_pixels: target_width, target_height = width, height else: scale_factor = math.sqrt(max_pixels / current_pixels) target_width = round(width * scale_factor) target_height = round(height * scale_factor) # Apply smart_resize for final dimensions final_height, final_width = smart_resize(target_height, target_width) return final_width, final_height def load_image(img_path: str) -> Image.Image: """Load image from URL or local path.""" if img_path.startswith("https://"): response = requests.get(img_path) return Image.open(BytesIO(response.content)) else: return Image.open(img_path) def visualize_points(original_image: Image.Image, points: list, new_width: int, new_height: int, original_width: int, original_height: int) -> None: """Draw prediction points on original image and save as output.png.""" output_img = original_image.copy() draw = ImageDraw.Draw(output_img) font = ImageFont.load_default(size=100) for i, point_data in enumerate(points): coords = point_data['point_2d'] # Map coordinates from resized image back to original image original_x = int(coords[0] / new_width * original_width) original_y = int(coords[1] / new_height * original_height) label = str(i + 1) # Draw circle circle_radius = 20 draw.ellipse([original_x - circle_radius, original_y - circle_radius, original_x + circle_radius, original_y + circle_radius], fill=(255, 0, 0)) # Draw label draw.text((original_x + 20, original_y - 20), label, fill=(255, 0, 0), font=font) print(f"Point {i+1}: Predicted coordinates {coords} -> Mapped coordinates [{original_x}, {original_y}]") output_img.save("output.png") print(f"Visualization with {len(points)} points saved to output.png") def main(): # Load model and processor model = Qwen2_5_VLForConditionalGeneration.from_pretrained( "InfiX-ai/InfiGUI-G1-7B", torch_dtype=torch.bfloat16, attn_implementation="flash_attention_2", device_map="auto" ) processor = AutoProcessor.from_pretrained("InfiX-ai/InfiGUI-G1-7B", padding_side="left") # Load and process image img_path = "https://raw.githubusercontent.com/InfiXAI/InfiGUI-G1/main/assets/test_image.png" image = load_image(img_path) # Store original image and resize for model input original_image = image.copy() original_width, original_height = image.size new_width, new_height = resize_image(original_width, original_height, MAX_IMAGE_PIXELS) resized_image = image.resize((new_width, new_height)) # Prepare model inputs instruction = "shuffle play the current playlist" system_prompt = 'You FIRST think about the reasoning process as an internal monologue and then provide the final answer.\nThe reasoning process MUST BE enclosed within tags.' prompt = f'''The screen's resolution is {new_width}x{new_height}. Locate the UI element(s) for "{instruction}", output the coordinates using JSON format: [{{"point_2d": [x, y]}}, ...]''' messages = [ {"role": "system", "content": system_prompt}, { "role": "user", "content": [ {"type": "image", "image": resized_image}, {"type": "text", "text": prompt} ] } ] # Generate predictions text = processor.apply_chat_template([messages], tokenize=False, add_generation_prompt=True) image_inputs, video_inputs = process_vision_info([messages]) inputs = processor(text=text, images=image_inputs, videos=video_inputs, padding=True, return_tensors="pt").to("cuda") generated_ids = model.generate(**inputs, max_new_tokens=512) output_text = processor.batch_decode( [out_ids[len(in_ids):] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)], skip_special_tokens=True, clean_up_tokenization_spaces=False ) # Parse and visualize results output_text = output_text[0].split("")[-1].replace("```json", "").replace("```", "").strip() output = json.loads(output_text) if output: visualize_points(original_image, output, new_width, new_height, original_width, original_height) if __name__ == "__main__": main() ``` ## Results Our InfiGUI-G1 models, trained with the AEPO framework, establish new state-of-the-art results among open-source models across a diverse and challenging set of GUI grounding benchmarks:

Model	MMBench-GUI	ScreenSpot-v2	UI-Vision	I2E-Bench	ScreenSpot-Pro
Qwen2.5-VL-7B	33.9	88.8	0.9	53.8	-
GUI-G²-7B	-	93.3	-	-	47.5
UI-TARS-7B	-	91.6	17.6	61.4	35.7
UGround-v1-7B	65.7	-	12.9	70.3	-
UI-TARS-1.5-7B	64.3	-	-	73.2	49.6
Qwen2.5-VL-72B	41.8	-	-	51.4	-
UGround-v1-72B	-	-	23.2	76.3	-
UI-TARS-72B	74.3	90.3	25.5	73.7	-
Ours
InfiGUI-G1-7B	80.8	93.5	26.1	77.4	51.9
w/ Expl. Success	86.4	95.6	34.4	83.0	58.0

## Evaluation This section provides instructions for reproducing the evaluation results reported in our paper. ### 1. Getting Started Clone the repository and navigate to the project directory: ```bash git clone https://github.com/InfiXAI/InfiGUI-G1.git cd InfiGUI-G1 ``` ### 2. Environment Setup The evaluation pipeline is built upon the [vLLM](https://github.com/vllm-project/vllm) library for efficient inference. For detailed installation guidance, please refer to the official vLLM repository. The specific versions used to obtain the results reported in our paper are as follows: - **Python**: `3.10.12` - **PyTorch**: `2.6.0` - **Transformers**: `4.50.1` - **vLLM**: `0.8.2` - **CUDA**: `12.6` The reported results were obtained on a server equipped with 4 x NVIDIA H800 GPUs. ### 3. Model Download Download the InfiGUI-G1 models from the Hugging Face Hub into the `./models` directory. ```bash # Create a directory for models mkdir -p ./models # Download InfiGUI-G1-3B huggingface-cli download --resume-download InfiX-ai/InfiGUI-G1-3B --local-dir ./models/InfiGUI-G1-3B # Download InfiGUI-G1-7B huggingface-cli download --resume-download InfiX-ai/InfiGUI-G1-7B --local-dir ./models/InfiGUI-G1-7B ``` ### 4. Dataset Download and Preparation Download the required evaluation benchmarks into the `./data` directory. ```bash # Create a directory for datasets mkdir -p ./data # Download benchmarks huggingface-cli download --repo-type dataset --resume-download likaixin/ScreenSpot-Pro --local-dir ./data/ScreenSpot-Pro huggingface-cli download --repo-type dataset --resume-download ServiceNow/ui-vision --local-dir ./data/ui-vision huggingface-cli download --repo-type dataset --resume-download OS-Copilot/ScreenSpot-v2 --local-dir ./data/ScreenSpot-v2 huggingface-cli download --repo-type dataset --resume-download OpenGVLab/MMBench-GUI --local-dir ./data/MMBench-GUI huggingface-cli download --repo-type dataset --resume-download vaundys/I2E-Bench --local-dir ./data/I2E-Bench ``` After downloading, some datasets require unzipping compressed image files. ```bash # Unzip images for ScreenSpot-v2 unzip ./data/ScreenSpot-v2/screenspotv2_image.zip -d ./data/ScreenSpot-v2/ # Unzip images for MMBench-GUI unzip ./data/MMBench-GUI/MMBench-GUI-OfflineImages.zip -d ./data/MMBench-GUI/ ``` ### 5. Running the Evaluation To run the evaluation, use the `eval/eval.py` script. You must specify the path to the model, the benchmark name, and the tensor parallel size. Here is an example command to evaluate the `InfiGUI-G1-3B` model on the `screenspot-pro` benchmark using 4 GPUs: ```bash python eval/eval.py \ ./models/InfiGUI-G1-3B \ --benchmark screenspot-pro \ --tensor-parallel 4 ``` - **`model_path`**: The first positional argument specifies the path to the downloaded model directory (e.g., `./models/InfiGUI-G1-3B`). - **`--benchmark`**: Specifies the benchmark to evaluate. Available options include `screenspot-pro`, `screenspot-v2`, `ui-vision`, `mmbench-gui`, and `i2e-bench`. - **`--tensor-parallel`**: Sets the tensor parallelism size, which should typically match the number of available GPUs. Evaluation results, including detailed logs and performance metrics, will be saved to the `./output/{model_name}/{benchmark}/` directory. ## Citation Information If you find this work useful, we would be grateful if you consider citing the following papers: ```bibtex @misc{liu2025infiguig1advancingguigrounding, title={InfiGUI-G1: Advancing GUI Grounding with Adaptive Exploration Policy Optimization}, author={Yuhang Liu and Zeyu Liu and Shuanghe Zhu and Pengxiang Li and Congkai Xie and Jiasheng Wang and Xueyu Hu and Xiaotian Han and Jianbo Yuan and Xinyao Wang and Shengyu Zhang and Hongxia Yang and Fei Wu}, year={2025}, eprint={2508.05731}, archivePrefix={arXiv}, primaryClass={cs.AI}, url={https://arxiv.org/abs/2508.05731}, } ``` ```bibtex @article{liu2025infigui, title={InfiGUI-R1: Advancing Multimodal GUI Agents from Reactive Actors to Deliberative Reasoners}, author={Liu, Yuhang and Li, Pengxiang and Xie, Congkai and Hu, Xavier and Han, Xiaotian and Zhang, Shengyu and Yang, Hongxia and Wu, Fei}, journal={arXiv preprint arXiv:2504.14239}, year={2025} } ``` ```bibtex @article{liu2025infiguiagent, title={InfiGUIAgent: A Multimodal Generalist GUI Agent with Native Reasoning and Reflection}, author={Liu, Yuhang and Li, Pengxiang and Wei, Zishu and Xie, Congkai and Hu, Xueyu and Xu, Xinchen and Zhang, Shengyu and Han, Xiaotian and Yang, Hongxia and Wu, Fei}, journal={arXiv preprint arXiv:2501.04575}, year={2025} } ``` ## Acknowledgements We would like to express our gratitude for the following open-source projects: [VERL](https://github.com/volcengine/verl), [Qwen2.5-VL](https://github.com/QwenLM/Qwen2.5-VL) and [vLLM](https://github.com/vllm-project/vllm).