Research map / Vision-language and multimodal / Human and embodied modeling
Harmonizing Visual Representations for Unified Multimodal Understanding and Generation
Size Wu, Wenwei Zhang, Lumin Xu, Sheng Jin, Zhonghua Wu, Qingyi Tao, Wentao Liu, Wei Li, Chen Change Loy
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 |
|---|---|---|---|
| use MAR encoder instead of VQGAN or VAE as shared visual encoder for understanding (ablation) | Harmon-0.5B · POPE | accuracy: 63.8 -> 86.1 | Improved |
| use MAR encoder instead of SigLIP for understanding (ablation) | Harmon-0.5B · MMMU | accuracy: 35.2 -> 33.3 | No clear effect |
| include stage I vision-language alignment (ablation) | Harmon-0.5B · MME-P | score: 1003 -> 1148 | Improved |
| include stage III high-quality fine-tuning (ablation) | Harmon-0.5B · MJHQ-30K | FID: 16.85 -> 6.08 | Improved |
| include stage II comprehensive multimodal training (ablation) | Harmon-0.5B · GenEval | overall: 0.65 -> 0.71 | Improved |
| increase stage III image resolution from 256 to 512 (ablation) | Harmon-0.5B · MJHQ-30K | FID: 9.91 -> 6.08 | Improved |
| increase stage III image resolution from 256 to 512 (ablation) | Harmon-0.5B · GenEval | overall: 0.68 -> 0.71 | Improved |
| add understanding loss when training generation (unified vs single-task) (ablation) | Harmon · GenEval | overall | Improved |
| co-train understanding and generation instead of understanding only (ablation) | Harmon · Understanding benchmarks | accuracy | No clear effect |
| use a separate SigLIP encoder for understanding (Janus-like decoupled design) instead of a shared MAR encoder (ablation) | Harmon · GenEval | overall | Hurt |
| use Harmon-1.5B with shared MAR encoder versus Janus-Pro-1.5B on understanding | Harmon-1.5B · MMBench | accuracy: 75.5 -> 65.5 | Hurt |
| use Harmon shared MAR encoder for text-to-image generation | Harmon-1.5B · GenEval | overall: 0.73 -> 0.76 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- Janus — Janus: Decoupling Visual Encoding for Unified Multimodal Understanding and Generation (CVPR 2025) · better on 2 of 2 shared results
- Show-o — Show-o: One Single Transformer to Unify Multimodal Understanding and Generation (ICLR 2025) · better on 1 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
- Janus: Decoupling Visual Encoding for Unified Multimodal Understanding and Generation (CVPR 2025 · similarity 0.91)
- Generation Enhances Understanding in Unified Multimodal Models via Multi-Representation Generation (ICML 2026 · similarity 0.90)
- UVU: Improving Multimodal Understanding via Vision-Language Unified Autoregressive Paradigm (CVPR 2026 · similarity 0.90)
- Vision as a Dialect: Unifying Visual Understanding and Generation via Text-Aligned Representations (NeurIPS 2025 · similarity 0.90)
- Tuna-2: Pixel Embeddings Beat Vision Encoders for Multimodal Understanding and Generation (NeurIPS 2026 · similarity 0.89)
- Learning to Generate via Understanding: Understanding-Driven Intrinsic Rewarding for Unified Multimodal Models (CVPR 2026 · similarity 0.89)
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.
