Papers
arxiv:2504.09925

FUSION: Fully Integration of Vision-Language Representations for Deep Cross-Modal Understanding

Published on Apr 14
· Submitted by starriver030515 on Apr 15
Authors:
,
,

Abstract

We introduce FUSION, a family of multimodal large language models (MLLMs) with a fully vision-language alignment and integration paradigm. Unlike existing methods that primarily rely on late-stage modality interaction during LLM decoding, our approach achieves deep, dynamic integration throughout the entire processing pipeline. To this end, we propose Text-Guided Unified Vision Encoding, incorporating textual information in vision encoding to achieve pixel-level integration. We further design Context-Aware Recursive Alignment Decoding that recursively aggregates visual features conditioned on textual context during decoding, enabling fine-grained, question-level semantic integration. To guide feature mapping and mitigate modality discrepancies, we develop Dual-Supervised Semantic Mapping Loss. Additionally, we construct a Synthesized Language-Driven Question-Answer (QA) dataset through a new data synthesis method, prioritizing high-quality QA pairs to optimize text-guided feature integration. Building on these foundations, we train FUSION at two scales-3B, 8B-and demonstrate that our full-modality integration approach significantly outperforms existing methods with only 630 vision tokens. Notably, FUSION 3B surpasses Cambrian-1 8B and Florence-VL 8B on most benchmarks. FUSION 3B continues to outperform Cambrian-1 8B even when limited to 300 vision tokens. Our ablation studies show that FUSION outperforms LLaVA-NeXT on over half of the benchmarks under same configuration without dynamic resolution, highlighting the effectiveness of our approach. We release our code, model weights, and dataset. https://github.com/starriver030515/FUSION

Community

Paper author Paper submitter
edited 8 days ago
Model # Vis Tok. MMB_EN MMB_CN VizWiz POPE MM-Vet MME_P MME_C Seed-Image HallB LLaVA_W MMStar MME-RW RWQA CV-Bench MMVP AI2D MathVista MMMU SQA TextVQA OCRBench ChartQA DocVQA
<=4B Model Comparison
Qwen2.5VL 3B - 79.1 78.1 - 85.9 61.4 1592.4 607.5 74.0 46.6 - 56.3 53.1 65.4 - - 81.4 61.2 51.2 79.3 - 82.8 84.0 93.93
InternVL2 4B - 78.5 73.9 - 84.6 50.5 1532.8 531.8 73.2 42.4 - 53.9 52.1 60.5 - - 79.0 58.5 48.3 96.0 74.7 78.4 81.5 89.2
DeepSeek-VL2-Tiny - 74.6 72.1 - - 52.5 1548.3 357.1 72.3 39.6 - 45.9 - 64.2 - - 71.6 53.6 40.7 - 80.7 80.5 81.0 86.9
MM1.5 3B - - - - 88.1 41.0 1478.4 319.6 72.4 - 73.0 - - 56.9 - - 65.7 44.4 37.1 85.8 76.5 65.7 74.2 87.5
Phi 3.5-Vision - 75.5 64.2 58.2 82.2 46.5 1473.4 412.1 69.9 53.3 68.8 49.0 - 53.5 69.3 67.7 77.4 - 43.3 89.0 61.1 59.8 72.0 75.9
Florence-VL 3B 576 71.6 60.8 59.1 88.3 51.0 1498.7 403.9 70.6 58.1 71.1 44.9 - 60.4 70.2 64.7 73.8 52.2 41.8 84.6 69.1 63.0 70.7 -
FUSION 3B (ours) 780 79.5 71.7 64.6 88.9 57.2 1595.9 416.5 74.6 51.4 84.7 52.4 41.5 65.1 76.4 76.0 78.9 54.3 44.7 87.1 71.8 60.0 75.7 70.9
FUSION-X 3B (ours) 620 80.3 74.8 66.1 88.7 60.3 1582.1 440.0 75.3 51.9 85.2 50.9 41.7 63.7 78.3 78.1 79.2 54.9 44.2 87.3 73.9 63.7 75.8 71.1
FUSION-L 3B (ours) 308 77.6 70.8 65.3 88.3 56.7 1573.7 406.8 74.1 48.7 77.6 44.7 39.5 61.8 76.2 77.0 77.3 48.6 43.4 85.6 71.4 56.9 67.7 63.5
>=7B Model Comparison
Qwen2VL 7B - 83.0 80.5 - 88.4 62.0 1639.2 637.1 76.0 50.6 - 60.7 57.4 70.1 - - 83.0 58.2 54.1 85.5 84.3 86.6 83.0 94.5
InternVL2 8B - 81.7 81.2 - 86.9 54.2 1639.7 575.3 75.4 45.2 - 61.5 53.5 64.4 - - 83.6 58.3 52.6 96.3 77.4 79.4 83.3 91.6
LLaVA-OneVision 8B - 81.7 78.0 - 87.2 58.8 1626.0 483.0 74.8 47.5 86.9 60.9 57.5 65.5 - - 81.6 56.1 47.7 96.6 78.5 69.7 78.8 87.5
MM1.5 7B - - - - 88.6 42.2 1514.9 346.4 73.4 - 74.2 - - 62.5 - - 72.2 47.6 41.8 89.6 76.5 63.5 88.1 78.2
Cambrian 8B 576 75.9 67.9 - 87.4 48.0 1547.1 - 74.7 48.7 71.0 50.0 - 64.2 72.2 51.3 73.0 49.0 42.7 80.4 71.7 62.4 73.3 77.8
Florence-VL 8B 576 76.2 69.5 59.1 89.9 56.3 1560.0 381.1 74.9 57.3 74.2 50.0 - 64.2 73.4 73.3 74.2 55.5 43.7 85.9 74.2 63.4 74.7 -
Eagle 8B 1024 75.9 - - - - 1559.0 - 76.3 - - - - 66.5 - 71.6 76.1 52.7 43.8 84.3 77.1 62.6 80.1 86.6
FUSION 8B (ours) 780 80.5 74.9 59.5 89.3 60.0 1592.3 396.1 77.2 52.6 86.9 52.4 46.0 65.2 78.7 78.7 80.4 56.6 43.1 89.2 77.3 63.8 80.3 78.6
FUSION-X 8B (ours) 620 82.0 76.2 62.9 88.8 60.0 1607.5 337.2 78.2 51.4 88.0 52.7 44.7 66.1 79.2 79.9 81.4 59.4 42.2 90.3 74.7 66.6 79.8 77.8
FUSION-L 8B (ours) 308 80.0 73.6 59.9 88.5 57.3 1601.7 338.9 75.9 46.7 82.1 49.3 42.3 65.1 78.2 76.7 79.2 55.2 41.8 88.3 72.8 59.5 73.0 66.0

With only 630 vision tokens, FUSION-X outperforms Cambrian-1 and Florence-VL, matching LLaVA-OneVision and nearly reaching the performance of top models like InternVL2 and Qwen2VL. Even with 300 vision tokens, FUSION-L retains 95% of its original performance, staying on par with Florence-VL.

Notably, FUSION-X 3B achieved the highest score on MMBench among models under 4B in size, even surpassing Qwen2.5VL 3B!

This is an automated message from the Librarian Bot. I found the following papers similar to this paper.

The following papers were recommended by the Semantic Scholar API

Please give a thumbs up to this comment if you found it helpful!

If you want recommendations for any Paper on Hugging Face checkout this Space

You can directly ask Librarian Bot for paper recommendations by tagging it in a comment: @librarian-bot recommend

Your need to confirm your account before you can post a new comment.

Sign up or log in to comment

Models citing this paper 4

Datasets citing this paper 2

Spaces citing this paper 0

No Space linking this paper

Cite arxiv.org/abs/2504.09925 in a Space README.md to link it from this page.

Collections including this paper 10