Research map / Vision-language and multimodal / Robot action and navigation
WeMMU: Enhanced Bridging of Vision-Language Models and Diffusion Models via Noisy Query Tokens
Jian Yang, Dacheng Yin, Xiaoxuan He, Yong Li, Fengyun Rao, Jing Lyu, Wei Zhai, Yang Cao, Zheng-Jun Zha
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 |
|---|---|---|---|
| make learnable queries dynamic by resolution (ablation) | WeMMU query bridge (Qwen2.5-VL-3B + Sana 1.6B) · ImageEdit-Bench (Stage 3) | Overall: 2.53 -> 2.88 | Improved |
| replace learnable queries with Noisy Query Tokens (ablation) | WeMMU query bridge (Qwen2.5-VL-3B + Sana 1.6B) · ImageEdit-Bench (Stage 3) | Overall: 2.88 -> 2.98 | Improved |
| add VAE branch with linear layer (ablation) | WeMMU query bridge (Qwen2.5-VL-3B + Sana 1.6B) · ImageEdit-Bench (Stage 3) | Overall: 2.98 -> 3.31 | Improved |
| add noisy queries vs learnable queries for multi-image editing (Stage 4) (ablation) | WeMMU query bridge (Qwen2.5-VL-3B + Sana 1.6B) · multi-image editing | qualitative success | Improved |
| add learnable channel-wise scaling to noise (ablation) | WeMMU query bridge (Qwen2.5-VL-3B + Sana 1.6B) · WeMMU | performance: scale mean 1.0, std 0.0074 | No clear effect |
| add Stage 4 multi-image training | WeMMU query bridge (Qwen2.5-VL-3B + Sana 1.6B) · GenEval | overall: 0.88 -> 0.88 | No clear effect |
| add Stage 4 multi-image training | WeMMU query bridge (Qwen2.5-VL-3B + Sana 1.6B) · ImageEdit-Bench | Overall: 3.31 -> 3.30 | No clear effect |
| unfreeze Qwen2.5-VL ViT (ablation) | WeMMU query bridge (Qwen2.5-VL-3B + Sana 1.6B) · WeMMU | training stability | Hurt |
| connect VAE branch with a linear layer instead of 2/6-layer ViT (ablation) | WeMMU query bridge (Qwen2.5-VL-3B + Sana 1.6B) · WeMMU | training loss convergence | Improved |
| use contrastive flow matching in later stages (ablation) | WeMMU query bridge (Qwen2.5-VL-3B + Sana 1.6B) · WeMMU | performance | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- OmniGen2 — OmniGen2: Towards Instruction-Aligned Multimodal Generation (CVPR 2026) · better on 2 of 5 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
- Accelerating Multimodal Large Language Models by Searching Optimal Vision Token Reduction (CVPR 2025 · similarity 0.89)
- Predictive Regularization Against Visual Representation Degradation in Multimodal Large Language Models (CVPR 2026 · similarity 0.89)
- LLaDA-V: Large Language Diffusion Models with Visual Instruction Tuning (CVPR 2026 · similarity 0.89)
- Vision‑Language‑Vision Auto‑Encoder: Scalable Knowledge Distillation from Diffusion Models (NeurIPS 2025 · similarity 0.89)
- Mitigating Mask Prior Drift and Positional Attention Collapse in Large Diffusion Vision-Language Models (ICML 2026 · similarity 0.89)
- Boosting Multimodal Large Language Models with Visual Tokens Withdrawal for Rapid Inference (AAAI 2025 · 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.
