Research map / Image generation / Text-to-image diffusion
Mixture of States: Routing Token-Level Dynamics for Multimodal Generation
Haozhe Liu, Ding Liu, Mingchen Zhuge, Zijian Zhou, Tian Xie, Sen He, Yukang Yang, Shuming Liu, Yuren Cong, Jiadong Guo, Hongyu Xu, Ke Xu and 7 more
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 |
|---|---|---|---|
| feed latent and timestep to the router in addition to the prompt (ablation) | MoS router · MJHQ | FID: 21.12 -> 20.15 | Improved |
| add noised latent to router input (without timestep) (ablation) | MoS router · MJHQ | FID: 21.12 -> 21.89 | Hurt |
| add latent and timestep to router input (ablation) | MoS router · MJHQ | CLIP score: 21.40 -> 21.74 | Improved |
| use token-specific routing instead of sample-wise routing (ablation) | MoS router · MJHQ ablation | FID and CLIP | Improved |
| use MoS adaptive routing instead of fixed-layer cross-attention (ablation) | MoS vs cross-attention · GenEval | overall: 0.74 -> 0.79 | Improved |
| use MoS adaptive routing instead of fixed-layer cross-attention (ablation) | MoS vs cross-attention · DPG-Bench | overall: 83.40 -> 85.61 | Improved |
| use MoS routing instead of MoT one-to-one layer correspondence (ablation) | MoS vs MoT (PLM-1B towers) · GenEval and DPG | score | Improved |
| add self-CoT reasoning caption | MoS-L · WISE | overall: 0.54 -> 0.65 | Improved |
| add self-CoT reasoning caption | MoS-S · WISE | overall: 0.47 -> 0.55 | Improved |
| use top-k=2 instead of k=1 or larger k (ablation) | MoS router · ablation (training curves) | performance | Improved |
| add epsilon-greedy exploration (ablation) | MoS router · training convergence | convergence speed | Improved |
| use separate RMSNorm per modality in the router (ablation) | MoS router · MJHQ FID/CLIP | FID and CLIP | Improved |
Closest papers by meaning
- Diffuse Everything: Multimodal Diffusion Models on Arbitrary State Spaces (ICML 2025 · similarity 0.90)
- ALTER: All-in-One Layer Pruning and Temporal Expert Routing for Efficient Diffusion Generation (NeurIPS 2025 · similarity 0.88)
- Faster Diffusion: Rethinking the Role of the Encoder for Diffusion Model Inference (NeurIPS 2024 · similarity 0.87)
- Scale Where It Matters: Training-Free Localized Scaling for Diffusion Models (NeurIPS 2026 · similarity 0.87)
- Diffusion Model Patching via Mixture-of-Prompts (AAAI 2025 · similarity 0.87)
- Warm Diffusion: Recipe for Blur-Noise Mixture Diffusion Models (ICLR 2025 · similarity 0.87)
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.
