Research map / Image generation / Fast diffusion sampling
MAMBO-G: Magnitude-Aware Mitigation for Boosted Guidance
Shangwen Zhu, Qianyu Peng, Zhilei Shu, Yuting Hu, Han Zhang, Andy Zheng, Xinyu Cui, Jian Zhao, Ruili Feng, Fan Cheng
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 ratio-based adaptive guidance damping (MAMBO-G) to CFG | Qwen-Image · 1024x1024 text-to-image | ImageReward: 0.20 -> 1.02 | Improved |
| add MAMBO-G to CFG at low resolution | Qwen-Image · 256x256 text-to-image | ImageReward: 0.53 -> 0.83 | Improved |
| add MAMBO-G to CFG with 10 steps | CFG · MS-COCO 5k prompts | FID: 63.62 -> 32.05 | Improved |
| use instance-level ratio damping instead of a fixed time-based schedule (ablation) | MAMBO-G · Qwen-Image 10 steps | ImageReward: 0.83 -> 1.08 | Improved |
| add time-based guidance schedule to constant CFG (ablation) | CFG · Qwen-Image 10 steps | ImageReward: 0.12 -> 0.83 | Improved |
| stack MAMBO-G on Guidance Rescale | Rescale · Qwen-Image | ImageReward: 0.73 -> 1.12 | Improved |
| stack MAMBO-G on APG | APG · Qwen-Image | ImageReward: 0.85 -> 0.96 | Improved |
| increase w_max from 6 to 16 (ablation) | MAMBO-G · Qwen-Image | ImageReward: 1.112 -> 0.780 | Hurt |
| increase w_max from 6 to 12 (ablation) | MAMBO-G · SD3.5 | ImageReward: 0.755 -> 0.878 (peak), 0.819 at 16 | Mixed |
| vary alpha from 6 to 16 (ablation) | MAMBO-G · Qwen-Image | ImageReward: 0.830 -> 1.161 (alpha 12) -> 1.122 (alpha 16) | Mixed |
| use exponential decay instead of linear or inverse mapping (ablation) | MAMBO-G · text-to-image | ImageReward | Improved |
| apply MAMBO-G on UniPC solver | MAMBO-G · SD3.5 UniPC | ImageReward | Improved |
Closest papers by meaning
- Improving Motion in Image-to-Video Models via Adaptive Low-Pass Guidance (CVPR 2026 · similarity 0.85)
- Eliminating Oversaturation and Artifacts of High Guidance Scales in Diffusion Models (ICLR 2025 · similarity 0.84)
- Visual Generation Without Guidance (ICML 2025 · similarity 0.83)
- Understanding and Improving Training-free Loss-based Diffusion Guidance (NeurIPS 2024 · similarity 0.83)
- Towards a Golden Classifier-Free Guidance Path via Foresight Fixed Point Iterations (NeurIPS 2025 · similarity 0.82)
- Accelerating Rectified Flow Models via Trajectory-Aware Caching (NeurIPS 2026 · similarity 0.82)
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.
