Research map / Vision-language and multimodal / Multimodal LLMs
LightFusion: A Light-weighted, Double Fusion Framework for Unified Multimodal Understanding and Generation
Zeyu Wang, Zilong Chen, Chenhui Gou, Feng Li, Chaorui Deng, Deyao Zhu, Kunchang Li, Weihao Yu, Haoqin Tu, Haoqi Fan, Cihang Xie
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add VAE tokens alongside ViT tokens for editing (ablation) | LightFusion · GEdit-EN | overall: 3.91 -> 5.61 | Improved |
| add VAE tokens alongside ViT tokens for editing (ablation) | LightFusion · ImgEdit | overall: 2.65 -> 3.57 | Improved |
| add ViT tokens alongside VAE tokens for editing (ablation) | LightFusion · GEdit-EN | overall: 4.93 -> 5.61 | Improved |
| add ViT tokens alongside VAE tokens for editing (ablation) | LightFusion · ImgEdit | overall: 3.38 -> 3.57 | Improved |
| raise training timestep shift from 1 to 4 (ablation) | LightFusion · DPG-Bench | overall: 76.67 -> 81.77 | Improved |
| raise training timestep shift from 1 to 4 (ablation) | LightFusion · ImgEdit | overall: 3.07 -> 3.57 | Improved |
| raise training timestep shift from 1 to 2 (ablation) | LightFusion · DPG-Bench | overall: 76.67 -> 78.84 | Improved |
| raise training timestep shift from 1 to 2 (ablation) | LightFusion · ImgEdit | overall: 3.07 -> 3.36 | Improved |
| condition DiT block i on VLM block i (0% depth deep fusion) instead of only final VLM output (ablation) | LightFusion · text-to-image and image editing | benchmark scores | Improved |
| hybrid ViT+VAE feature fusion rather than ViT-only conditioning (vs UniWorld-V1) | LightFusion · GEdit-Bench-EN | semantic consistency: 4.93 -> 6.34 | Improved |
| deep Double Fusion model vs UniWorld-V1 | LightFusion · ImgEdit-Bench | overall: 3.26 -> 3.77 | Improved |
| Double Fusion with frozen understanding branch | LightFusion · MMBench/MMMU/MM-Vet | accuracy: 83.5, 58.6, 67.1 | No clear effect |
Later papers that compared with this method
1 later accepted paper reports LightFusion as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- VQ-VA World: Towards High-Quality Visual Question-Visual Answering (CVPR 2026 · better on 8 of 8 shared results)
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 2 on most shared results.
- OmniGen2 — OmniGen2: Towards Instruction-Aligned Multimodal Generation (CVPR 2026) · better on 2 of 2 shared results
- Show-o2 — Show-o2: Improved Native Unified Multimodal Models (NeurIPS 2025) · better on 0 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- DuoGen: Towards Autonomous Interleaved Multimodal Generation (CVPR 2026 · similarity 0.89)
- DenseFusion-1M: Merging Vision Experts for Comprehensive Multimodal Perception (NeurIPS 2024 · similarity 0.89)
- LMFusion: Adapting Pretrained Language Models for Multimodal Generation (NeurIPS 2025 · similarity 0.89)
- MANZANO: A Simple and Scalable Unified Multimodal Model with a Hybrid Vision Tokenizer (ICLR 2026 · similarity 0.89)
- UNICOM: Unified Multimodal Modeling via Compressed Continuous Semantic Representations (NeurIPS 2026 · similarity 0.88)
- The Narrow Gate: Localized Image-Text Communication in Native Multimodal Models (NeurIPS 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
