---
license: apache-2.0
language:
- en
base_model:
- microsoft/Phi-4-mini-instruct
- facebook/dinov2-with-registers-giant
- google/siglip2-so400m-patch14-224
base_model_relation: adapter
pipeline_tag: image-text-to-text
---
# Aurea: Adaptive Multimodal Fusion for Vision-Language Models
Aurea is an open-source research framework centered on an adaptive spatial-range attention module that fuses spatial and semantic cues from encoder features, yielding richer, context-aware representations for downstream tasks.
[Explore the full source code and technical documentation on GitHub](https://github.com/Dcas89/Aurea)
## Key Features
- **Multiple Vision Encoders:** Input images are encoded separately by DINOv2 and SigLIP2.
- **Multi-stage Fusion:** The `SpatialRangeBlock` fuses these inputs through multiple layers of `SpatialRangeAttention`, which selectively aggregates features by jointly considering spatial proximity and semantic similarity. This is performed with a highly optimized fused CUDA kernel.
- **Flexible Language Model Integration:** While Phi-4 is the default language model, Aurea is designed for easy adaptation to other pretrained language models with minimal engineering effort.
- **Model Weights:** Two model checkpoints are provided: (1) base pretrained weights (trained on a ~558k image subset of LAION) and (2) instruction-tuned weights (further fine-tuned on ~625k samples from LLaVA 1.5 datasets). All checkpoints can be downloaded directly from this repository.
- **Extensible and Modular:** The code supports straightforward extension, experimentation, and integration with novel encoders or downstream tasks.
## Installation
1. **Clone the source repository**
```bash
git clone https://github.com/Dcas89/Aurea.git
cd Aurea
```
2. **Install Python dependencies**
```bash
pip install -r requirements.txt
```
## Usage
First, initialize the Aurea model:
```python
from entry import Aurea
aurea = Aurea(root_dir='/path/to/Aurea')
```
> **Note:** When initializing the model, all required model checkpoints will be downloaded automatically.
### Image + Text Generation (Basic)
Generate text based on an image and prompt:
```python
# Basic image + text generation
response = aurea.generate(
prompt="How many remote control devices are in this image?",
image_path='./assets/cats.png' # Example image included in the repo
)
print(response)
```
### Generation with Custom Parameters
Tune generation parameters for more control:
```python
# Advanced generation with custom parameters
response = aurea.generate(
prompt="Only one cat is wearing a collar in the image. Which cat is it? Answer Briefly: Left, Right, or Both",
image_path='./assets/cats.png', # Example image included in the repo
max_new_tokens=50, # Maximum number of tokens to generate
temperature=0.1, # Lower values make output more deterministic
repetition_penalty=1.1, # Penalizes token repetition (>1.0)
filter_kwargs={'thres': 0.90, 'top_k': 50}, # Parameters for filtering function
use_dynamic_top_k=False, # Whether to use dynamic top-k sampling
min_top_k=50, # Minimum top-k value if using dynamic top-k
max_top_k=90, # Maximum top-k value if using dynamic top-k
filter_fn=None, # Custom filtering function
exclude_prompt=True # Whether to exclude prompt from returned text
)
print(response)
```
### Logit Filtering
Using a specific filtering function (e.g., top_p):
```python
from generate import top_p
response = aurea.generate(
prompt="Only one cat is wearing a collar in the image. What is the color of the collar? Answer Briefly: Blue, Light Green, Yellow",
image_path='./assets/cats.png', # Example image included in the repo
max_new_tokens=50,
temperature=0.1,
repetition_penalty=1.1,
filter_kwargs={'thres': 0.99, 'top_k': 50},
filter_fn=top_p, # Using top-p sampling
exclude_prompt=True
)
print(response)
```
### Dynamic Top-K Sampling
Example using dynamic top-k sampling (interpolating from max_top_k to min_top_k over generation):
```python
response = aurea.generate(
prompt="What does the logo say and what does it represent?",
image_path='./assets/mazure.png',
max_new_tokens=100,
temperature=0.1,
repetition_penalty=1.1,
filter_kwargs={'thres': 0.99, 'top_k': 50},
use_dynamic_top_k=True, # Enable dynamic top-k sampling
min_top_k=50, # Lower bound for top-k
max_top_k=90, # Upper bound for top-k
filter_fn=None,
exclude_prompt=True
)
print(response)
```
### Text-Only Generation
Aurea can also be used for text-only tasks:
```python
# Text-only generation (no image)
response = aurea.generate(
prompt="What is CUDA programming?",
max_new_tokens=200,
temperature=0.1,
repetition_penalty=1.1,
filter_kwargs={'thres': 0.9, 'top_k': 50},
exclude_prompt=True
)
print(response)
```
## References
- [SigLIP 2: Multilingual Vision-Language Encoders](https://doi.org/10.48550/arXiv.2502.14786)
- [Phi-4 Technical Report](https://doi.org/10.48550/arXiv.2412.08905)
- [DINOv2: Learning Robust Visual Features without Supervision](https://doi.org/10.48550/arXiv.2304.07193)
- [LLaVA](https://github.com/haotian-liu/LLaVA)
- [LLaVA-UHD](https://github.com/thunlp/LLaVA-UHD)
## License
This project is released under the Apache 2.0 License.
## Acknowledgements
- The CUDA spatial-range attention is inspired by and adapted from LLaVA-UHD.
- Some components were adapted from [lucidrains](https://github.com/lucidrains) repositories, which provide excellent implementations of various transformer and attention mechanisms.
- Thanks to the open-source community for DINOv2, SigLIP2, LLaVA, LlaVA-UHD, and Phi-4.
- Thanks to Hugging Face for their [Transformers](https://github.com/huggingface/transformers) and [Accelerate](https://github.com/huggingface/accelerate) libraries.
This project incorporates code and models from:
- Phi-4 Mini: Copyright (c) 2025 Microsoft Corporation
- DINOv2: Copyright (c) 2024 Meta Platforms, Inc.
- SigLIP2: Copyright (c) 2025 Google LLC