Research map / Vision-language and multimodal / Multimodal LLMs
MetaMorph: Multimodal Understanding and Generation via Instruction Tuning
Shengbang Tong, David Fan, Jiachen Li, Yunyang Xiong, Xinlei Chen, Koustuv Sinha, Michael Rabbat, Yann LeCun, Saining Xie, Zhuang Liu
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 visual understanding data to joint training (vs generation-only) (ablation) | VPiT LLaMA-3 8B · COCO-30K, 200k generation data | FID: 110.7 -> 15.2 | Improved |
| add visual understanding data to joint training (vs generation-only) (ablation) | VPiT LLaMA-3 8B · COCO-30K, 5M generation data | FID: 27.7 -> 14.3 | Improved |
| joint train with ImageQA data (ablation) | VPiT · COCO-30K, 200k generation data | FID: 110.5 -> 18.9 | Improved |
| joint train with VideoQA data (ablation) | VPiT · COCO-30K, 200k generation data | FID: 110.5 -> 26.5 | Improved |
| joint train with pure video data (ablation) | VPiT · COCO-30K, 200k generation data | FID: 110.5 -> 84.7 | Improved |
| joint train with image-to-image data (ablation) | VPiT · COCO-30K, 200k generation data | FID: 110.5 -> 97.5 | Improved |
| increase VQA data from 1M to 7M (200k generation data) (ablation) | VPiT · Image QA average | accuracy: 46.4 -> 55.8 | Improved |
| increase VQA data from 1M to 7M (200k generation data) (ablation) | VPiT · COCO-30K | FID: 28.3 -> 18.2 | Improved |
| increase generation data from 200k to 4M (1M VQA) (ablation) | VPiT · Image QA average | accuracy: 46.4 -> 51.4 | Improved |
| increase generation data from 200k to 4M (7M VQA) (ablation) | VPiT · Image QA average | accuracy: 55.8 -> 56.2 | No clear effect |
| increase generation data from 200k to 4M (1M VQA) (ablation) | VPiT · COCO-30K | FID: 28.3 -> 18.4 | Improved |
| use cosine similarity loss instead of L1 loss (ablation) | VPiT LLaMA-3 8B · Image QA average | accuracy: 53.83 -> 55.93 | Improved |
Later papers that compared with this method
2 later accepted papers report MetaMorph as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- LMFusion: Adapting Pretrained Language Models for Multimodal Generation (NeurIPS 2025 · better on 1 of 2 shared results)
- X-Fusion: Introducing New Modality to Frozen Large Language Models (ICCV 2025 · better on 1 of 1 shared result)
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.
- Janus — Janus: Decoupling Visual Encoding for Unified Multimodal Understanding and Generation (CVPR 2025) · better on 1 of 2 shared results
- VILA-U-256 — VILA-U: a Unified Foundation Model Integrating Visual Understanding and Generation (ICLR 2025) · better on 2 of 2 shared results
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
- Learning to Instruct for Visual Instruction Tuning (NeurIPS 2025 · similarity 0.83)
- Pilot: Building the Federated Multimodal Instruction Tuning Framework (AAAI 2025 · similarity 0.83)
- INST-IT: Boosting Instance Understanding via Explicit Visual Prompt Instruction Tuning (NeurIPS 2025 · similarity 0.83)
- OmniGen2: Towards Instruction-Aligned Multimodal Generation (CVPR 2026 · similarity 0.83)
- Re-Imagining Multimodal Instruction Tuning: A Representation View (ICLR 2025 · similarity 0.83)
- CoDi-2: In-Context Interleaved and Interactive Any-to-Any Generation (CVPR 2024 · similarity 0.83)
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.
