README.md

---
license: mit
base_model:
- microsoft/Phi-3.5-vision-instruct
tags:
- GUI
- Agent
- Grounding
- CUA
---

# Microsoft Phi-Ground-Any-4B

<p align="center">
   <a href="https://microsoft.github.io/Phi-Ground/" target="_blank">🤖 HomePage</a> | <a href="https://arxiv.org/abs/2605.12501" target="_blank">📄 Paper </a> | <a href="https://arxiv.org/abs/2605.12501" target="_blank">📄 Arxiv </a> | <a href="https://huggingface.co/microsoft/Phi-Ground-Any" target="_blank"> 😊 Model </a> | <a href="https://github.com/microsoft/Phi-Ground/tree/main/benchmark/" target="_blank"> 😊 Eval data </a> 
</p>

![overview](docs/images/intro.png)

**Phi-Ground-Any-4B** is one of the Phi-Ground model family, finetuned from [microsoft/Phi-3.5-vision-instruct](https://huggingface.co/microsoft/Phi-3.5-vision-instruct) with fixed input resolution 1680x1008. 

### Main results

![overview](docs/images/r1.png)

### Usage
The current `transformers` version can be verified with: `pip list | grep transformers`.

Examples of required packages:
```
flash_attn==2.5.8
numpy==1.24.4
Pillow==10.3.0
Requests==2.31.0
torch==2.3.0
torchvision==0.18.0
transformers==4.43.0
accelerate==0.30.0
```


### Input Formats

The model require strict input format including fixed image resolution, instruction-first order and system prompt.

Input preprocessing

```python
from PIL import Image


def process_image(img):
    # Phi-Ground-Anything uses a larger 5x3-tile canvas (1680 x 1008).
    target_width, target_height = 336 * 5, 336 * 3

    img_ratio = img.width / img.height
    target_ratio = target_width / target_height

    if img_ratio > target_ratio:
        new_width = target_width
        new_height = int(new_width / img_ratio)
    else:
        new_height = target_height
        new_width = int(new_height * img_ratio)
    reshape_ratio = new_width / img.width

    img = img.resize((new_width, new_height), Image.LANCZOS)
    new_img = Image.new("RGB", (target_width, target_height), (255, 255, 255))
    paste_position = (0, 0)
    new_img.paste(img, paste_position)
    return new_img, reshape_ratio


# Phi-Ground-Anything takes the user instruction directly (no "describe the
# element" wrapper) and is trained to emit the click point as
#   <x>VALUE</x><y>VALUE</y>
# where VALUE is a relative coordinate in [0, 10000] over the padded canvas
# (i.e., divide by 10000 and multiply by target_width / target_height to get
# pixel coords in the padded image, then divide by reshape_ratio to recover
# coords in the ORIGINAL image).
instruction = "<your instruction>"
prompt = """<|user|> 
{instruction}<|image_1|> 
<|end|> 
<|assistant|>""".format(instruction=instruction)

image_path = "<your image path>"
original_image = Image.open(image_path).convert("RGB")
image, reshape_ratio = process_image(original_image)


# ---------------------------------------------------------------------------
# Example: parse the model output and recover original-image coordinates.
# ---------------------------------------------------------------------------
import re

target_width, target_height = 336 * 5, 336 * 3
SCALE = 10000.0

x_pattern = re.compile(r"<x>\s*(-?\d+(?:\.\d+)?)\s*</x>")
y_pattern = re.compile(r"<y>\s*(-?\d+(?:\.\d+)?)\s*</y>")


def parse_xy(model_output: str):
    xs = [float(v) for v in x_pattern.findall(model_output)]
    ys = [float(v) for v in y_pattern.findall(model_output)]
    return list(zip(xs, ys))


def to_original_pixel(rel_xy, reshape_ratio: float):
    x_rel, y_rel = rel_xy
    px = (x_rel / SCALE) * target_width / reshape_ratio
    py = (y_rel / SCALE) * target_height / reshape_ratio
    return px, py


# model_output = "<x>4823</x><y>3120</y>"
# point_orig = to_original_pixel(parse_xy(model_output)[0], reshape_ratio)

```


Then you can use huggingface model or [vllm](https://github.com/vllm-project/vllm) to inference.