Research map / Image generation / Autoregressive and latent generation
MUSE: Unlocking Timestep as Native Task Steering for One-Step Dense Prediction
Shuo Zhou, Zhaoxin Li, Xiujuan Chai
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 |
|---|---|---|---|
| assign distinct timesteps to depth and normal tasks instead of an identical timestep (ablation) | MUSE-g · NYUv2 depth | AbsRel (%): 26.05 -> 9.67 | Improved |
| use high timesteps near 999 instead of 599 for all tasks (ablation) | MUSE-g · NYUv2 depth | AbsRel (%): 27.05 -> 8.95 | Improved |
| vary absolute timestep magnitude in the discriminative variant (ablation) | MUSE-d · NYUv2 depth | AbsRel (%): 6.31, 6.40, 6.48 | No clear effect |
| use native sinusoidal timestep embedding instead of Fourier embedding (ablation) | MUSE-g · NYUv2 depth | AbsRel (%): 10.02 -> 8.95 | Improved |
| use timestep key instead of learnable task token (ablation) | MUSE-g · NYUv2 depth | AbsRel (%): 12.76 -> 8.95 | Improved |
| use timestep key instead of learnable task token (ablation) | MUSE-g · iBims-1 normals | mean angular error (deg): 20.83 -> 19.67 | Improved |
| train depth and normal jointly instead of depth only (ablation) | MUSE-g · NYUv2 depth | AbsRel (%): 10.70 -> 11.07 | No clear effect |
| train depth and normal jointly instead of normal only (ablation) | MUSE-g · iBims-1 normals | mean angular error (deg): 20.01 -> 19.66 | No clear effect |
| steer with timesteps 995 and 999 instead of unsteered timesteps (ablation) | MUSE-g · Latent clustering | Silhouette score (last layer): -0.0166 -> 0.1572 | Improved |
| steer with timesteps 995 and 999 instead of unsteered timesteps (ablation) | MUSE-g · Latent clustering | Calinski-Harabasz (last layer): 0.89 -> 80.27 | Improved |
| remove injected noise (discriminative instead of generative variant) | MUSE · NYUv2 depth | AbsRel (%): 5.8 -> 5.1 | Improved |
| remove injected noise (discriminative instead of generative variant) | MUSE · KITTI depth | AbsRel (%): 9.8 -> 9.2 | Improved |
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.
- FE2E — FE2E: From Editor to Dense Geometry Estimator (CVPR 2026) · better on 0 of 4 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
- What Matters When Repurposing Diffusion Models for General Dense Perception Tasks? (ICLR 2025 · similarity 0.86)
- Adaptive Non-Uniform Timestep Sampling for Accelerating Diffusion Model Training (CVPR 2025 · similarity 0.85)
- Coarse-to-Fine Compositional Diffusion for Long-Horizon Planning (NeurIPS 2026 · similarity 0.85)
- Going Beyond Multi-Task Dense Prediction with Synergy Embedding Models (CVPR 2024 · similarity 0.85)
- Efficient Multitask Dense Predictor via Binarization (CVPR 2024 · similarity 0.85)
- Improving Training Efficiency of Diffusion Models via Multi-Stage Framework and Tailored Multi-Decoder Architecture (CVPR 2024 · similarity 0.85)
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.
