---
license: apache-2.0
language:
- en
pipeline_tag: image-text-to-text
---
# Model description
`xGen-MM` is a series of the latest foundational Large Multimodal Models (LMMs) developed by Salesforce AI Research. This series advances upon the successful designs of the `BLIP` series, incorporating fundamental enhancements that ensure a more robust and superior foundation. These models have been trained at scale on high-quality image caption datasets and interleaved image-text data.
In the v1.5 (08/2024) release, we present a series of XGen-MM models including:
- [🤗 xGen-MM-instruct-interleave (our main instruct model)](https://huggingface.co/Salesforce/xgen-mm-phi3-mini-instruct-multi-r-v1.5): `xgen-mm-phi3-mini-instruct-interleave-r-v1.5`
- This model has higher overall scores than [xGen-MM-instruct](https://huggingface.co/Salesforce/xgen-mm-phi3-mini-instruct-singleimg-r-v1.5) on both single-image and multi-image benchmarks.
- [🤗 xGen-MM-base](https://huggingface.co/Salesforce/xgen-mm-phi3-mini-base-r-v1.5): `xgen-mm-phi3-mini-base-r-v1.5`
- [🤗 xGen-MM-instruct](https://huggingface.co/Salesforce/xgen-mm-phi3-mini-instruct-singleimg-r-v1.5): `xgen-mm-phi3-mini-instruct-singleimg-r-v1.5`
- [🤗 xGen-MM-instruct-dpo](https://huggingface.co/Salesforce/xgen-mm-phi3-mini-instruct-dpo-r-v1.5): `xgen-mm-phi3-mini-instruct-dpo-r-v1.5`
In addition to the models, our team also released a series of datasets for multi-modal pre-training, including:
- [🍃 MINT-1T: Scaling Open-Source Multimodal Data by 10x: A Multimodal Dataset with One Trillion Tokens](https://arxiv.org/abs/2406.11271)
- [🤗 BLIP3-OCR-200M](https://huggingface.co/datasets/Salesforce/blip3-ocr-200m): a dataset with dense OCR annotations.
- [🤗 BLIP3-GROUNDING-50M](https://huggingface.co/datasets/Salesforce/blip3-grounding-50m): a dataset for enhancing the ability to ground semantic concepts in images.
- BLIP3-KALE (stay tuned): a large-scale curated high-quality caption dataset.
For more details, check out our [tech report](https://arxiv.org/pdf/2408.08872), [fine-tuning code](https://github.com/salesforce/LAVIS/tree/xgen-mm), and project page (coming soon).
# Data
The instruct model is fine-tuned on a mixture of around 1 million samples from multiple domains. All the fine-tuning data are from public sources, most of which are covered in [The Cauldron](https://huggingface.co/datasets/HuggingFaceM4/the_cauldron).
# Results
### Single-image benchmarks
| Model (Size) | SEED -IMG | SEED v2 | MMB (dev) | MM Star | MME (norm) | CVB -2D | CVB -3D | RealW QA | MMMU (val) | Math Vista | Sci QA | POPE | Text VQA | Avg. all | Avg. perc. |
|--------------------------------|:---------:|:-------:|:----------:|:-------:|:-----------:|:-------:|:-----------------:|-------------------|:-----------------:|:-----------------:|:-----------------:|:-----------------:|----------------|:--------------:|----------------|
| Closed-source models | | | | | | | | | | | | | | | |
| GPT-4V* | 72.0 | - | 80.8 | 49.7 | 63.3 | 64.3 | 73.8 | 56.5 | 53.8 | 48.2 | 82.1 | 75.4 | - | - | - |
| MM1-3B-Chat (3B) | 68.8 | - | 67.8 | - | 62.9 | - | - | - | 33.9 | - | - | 87.4 | - | - | - |
| Open-source models | | | | | | | | | | | | | | | |
| HPT-1.5-edge (4B) | **72.3** | - | 74.6 | 45.8 | - | - | - | - | 42.6 | **45.1** | 85.4 | **91.0** | - | - | - |
| VILA-1.5-3B (3B) | 67.9 | - | 63.4 | - | - | - | - | - | 33.3 | - | 69.0 | 85.9 | - | - | - |
| VILA-1.5-3B** (3B) | 67.9 | 51.9 | 62.4 | 40.3 | 58.5 | 50.1 | 60.3 | 53.3 | 34.1 | 30.6 | 68.9 | 86.9 | 58.1 | 55.6 | 59.1 |
| phi-3-vision (4B) | - | - | 80.5 | - | - | - | - | - | - | 44.5 | 90.8 | 85.8 | 70.9 | - | - |
| phi-3-vision** (4B) | 71.0 | 52.7 | 74.2 | 47.9 | 55.3 | 60.7 | 68.2 | 59.1 | **46.1** | **45.1** | **90.2** | 83.5 | **73.3** | 63.6 | 63.6 |
| xGen-MM-inst. (4B) | 71.8 | 53.9 | 76 | 46.7 | 63.8 | 66.2 | **75.4** | **61.6** | 42.8 | 39.2 | 85.6 | 87.0 | 72.0 | 64.8 | 66.9 |
| **xGen-MM-inst.-interleave (4B)** | 72.2 | **55.5** | **76.8** | **48.1** | **64.4** | **69.3** | 72.3 | 60.5 | 41.1 | 39.6 | 88.3 | 87.0 | 71.0 | **65.1** | **67.3** |
* GPT-4V(gpt-4-1106-preview) results are taken from this third-party [leaderborad](https://huggingface.co/spaces/opencompass/open_vlm_leaderboard).
** Model results are tested with our evaluation code for a fair comparison.
### Multi-image benchmarks
| Model | BLINK | QBench-2 | Mantis-eval |
|-------------------------------|:-----------------:|:-----------------:|:-----------------:|
| GPT-4V † | 51.1 | 73.4 | 62.7 |
| VILA-1.5-3B†† (3B) | 39.8 | 51.7 | 41.9 |
| xGen-MM-inst. (4B) | 46.6 | 52.4 | 42.4 |
| **xGen-MM-inst.-interleave (4B)** | 49.7 | 75.1 | 56.7 |
† GPT-4V results are the numbers reported in each benchmark's original paper.
†† Model results are tested with our evaluation code for a fair comparison.
### Examples
# How to use
Please check out our [inference notebook](demo.ipynb) for example code to use our model. We also provide an example script for [batch inference](batch_inference.ipynb).
# Reproducibility:
Our evaluation is implemented based on [open-compass/VLMEvalKit](https://github.com/open-compass/VLMEvalKit). We will create a PR to that repo to support XGen-MM evaluation.
# Bias, Risks, Limitations, and Ethical Considerations
The main data sources are from the internet, including webpages,
image stock sites, and curated datasets released by the research community. We have excluded certain data, such as LAION, due to known CSAM concerns.
The model may be subject to bias from the original data source, as well as bias from LLMs and commercial APIs.
We strongly recommend users assess safety and fairness before applying to downstream applications.
# License
Our code and weights are released under the [Apache 2.0](https://www.apache.org/licenses/LICENSE-2.0.txt) license.
# Code acknowledgment
Our training code is based on [OpenFlamingo: An open-source framework for training large multimodal models.](https://github.com/mlfoundations/open_flamingo), and part of our data preprocessing code is adapted from [LLaVA](https://github.com/haotian-liu/LLaVA).
The evaluation code for the instruct models is based on [VLMEvalKit: Open-source evaluation toolkit of large vision-language models (LVLMs)](https://github.com/open-compass/VLMEvalKit).
We thank the authors for their open-source implementations.
# Citation
```
@article{blip3-xgenmm,
author = {Le Xue, Manli Shu, Anas Awadalla, Jun Wang, An Yan, Senthil Purushwalkam, Honglu Zhou, Viraj Prabhu, Yutong Dai, Michael S Ryoo, Shrikant Kendre, Jieyu Zhang, Can Qin, Shu Zhang, Chia-Chih Chen, Ning Yu, Juntao Tan, Tulika Manoj Awalgaonkar, Shelby Heinecke, Huan Wang, Yejin Choi, Ludwig Schmidt, Zeyuan Chen, Silvio Savarese, Juan Carlos Niebles, Caiming Xiong, Ran Xu},
title = {xGen-MM(BLIP-3): A Family of Open Large Multimodal Models},
journal = {arXiv preprint},
month = {August},
year = {2024},
}
```
# Troubleshoot
1. If you missed any packages, please consider the following
```
pip install torch==2.2.1 torchvision==0.17.1 torchaudio==2.2.1 --index-url https://download.pytorch.org/whl/cu121
pip install open_clip_torch==2.24.0
pip install einops
pip install einops-exts
pip install transformers==4.41.1
```
