Mantis: Interleaved Multi-Image Instruction Tuning

Balancing Multi-Image and Single-Image Abilities of Large Multimodal Models

Dongfu Jiang, Xuan He, Huaye Zeng, Cong Wei, Max Ku, Qian Liu, Wenhu Chen
▶ University of Waterloo ▶ Tsinghua University ▶ Sea AI Lab
Paper Code Mantis-Instruct Mantis-Eval 🔥 Models 🤗 Spaces Twitter

Abstract

🤔 The recent years have witnessed a great array of large multimodal models (LMMs) to effectively solve single-image vision language tasks. However, their abilities to solve multi-image visual language tasks is yet to be improved.
😦 The existing multi-image LMMs (e.g. OpenFlamingo, Emu, Idefics, etc) mostly gain their multi-image ability through pre-training on hundreds of millions of noisy interleaved image-text data from web, which is neither efficient nor effective.

  1. Mantis-Instruct Data. We present the first fully text-image interleaved multimodal instruction tuning dataset, containing 721K examples from 14 subsets and covering multi-image skills including co-reference, reasoning, comparing, temporal understanding
  2. Mantis Model. We introduce Mantis, an LLaMA-3 based LMM with interleaved text and image as inputs, train on Mantis-Instruct under academic-level resources (i.e. 36 hours on 16xA100-40G)
  3. Performance. Mantis reaches the state-of-the-art performance on five multi-image benchmarks (NLVR2, Q-Bench, BLINK, MVBench, Mantis-Eval), and also maintain a strong single-image performance on par with CogVLM and Emu2.
  4. Open-source. Our curated Mantis-Instruct data, along with the training/evaluation codes, model checkpoints are all released to the public

Interleaved Multi-Image Instruction-Tuning data

  • Mantis-Instruct has a total of 721K instances, consisting of 14 subsets to cover all the multi-image skills.
  • Among the 14 subsets, 10 subsets are from the existing datasets. For example, NLVR2, IconQA, etc for reasoning skill; DreamSim, Birds-to-Words, etc for comparison skill; NExT-QA, STAR, for temporal understanding
  • We additionally curate four new datasets LLaVA-665k-multi, LRV-multi to cover coref skill and Contrast-Caption, Multi-VQA to broaden reasoning skill, where Multi-VQA is generated by prompting GPT-4.
  • Please check out 🤗 Mantis-Instruct on hugging face datasets for usage.
Subset Multi-Image Skill Sample Size
LLaVA-665k-multi Coref 313K
LRV-multi Coref 8K
NLVR2 Reason 86K
IconQA Reason 64K
Contrast-Caption Reason 36K
ImageCoDe Reason 17K
Multi-VQA Reason 5K
Co-Instruct Compare 151K
Dreamsim Compare 16K
Spot-the-Diff Compare 8K
Birds-to-Words Compare 3K
VIST Temporal 7K
NExT-QA Temporal 4K
STAR Temporal 3K

Mantis:

Mantis applies the LLaVA's architecture, using CLIP/SigLIP as vision encoders and Meta-Llama-3-8B-Instruct as language model. We consider a two-stage instruction-tuning procedure: connects pre-trained CLIP ViT-L/14 visual encoder and large language model Vicuna, using a simple projection matrix. To support super-resolution, we also train a variant based on SigLIP and Fuyu-8B. We also train a variant Mantis-8B-Idefics2 based on idefics2-8b, which is the most competitive model of Mantis family. Please check out our [Model Zoo].

Performance

Multi-Image VQA: Towards GPT-4-level multi-image understanding

Benchmark Multi-Image Skill Held-in/Held-out
NLVR2 Reason Held-in
Q-bench Reason Held-in
Mantis-Eval Reason & Co-reference Held-out
BLINK Reason Held-out
MVBench Temporal Held-out

We select 5 multi-image benchmarks that cover the four crucial multi-image skills: co-reference, reasoning, comparing, temporal understanding to evaluate Mantis.

Evaluation on 5 benchmarks, including NLVR2, Q-Bench, BLINK, MVBench, Mantis-Eval, shows that Mantis achieves the state-of-the-art performance. It demonstrates that Mantis effectively learn the 4 crucial multi-image skills (co-reference, reasoning, comparing, temporal understanding) from the interleaved text-image instructions dataset, Mantis-Instruct. We have surpassed the second best Idefics2-8B (pre-trained on 140M interleaved image-text data) by an average of 9 absolute points, and is only behind GPT-4 by 2 points.

Single Image VQA: Maintain strong performance

We also evaluate Mantis-8B-CLIP and Mantis-8B-SigLIP on various single-image tasks, including TextVQA, VQA-v2, MMBench, MMMU, etc. Mantis models reach on-par average performance with CogVLM, and Emu2-Chat

Examples on Visual Instruction Following

BibTeX

@inproceedings{Jiang2024MANTISIM,
  title={MANTIS: Interleaved Multi-Image Instruction Tuning},
  author={Dongfu Jiang and Xuan He and Huaye Zeng and Cong Wei and Max W.F. Ku and Qian Liu and Wenhu Chen},
  publisher={arXiv2405.01483}
  year={2024},
}