Research map / Image generation / Fast diffusion sampling
Letting Trajectories Spread: Quality-Preserving Control for Diverse Flow Matching
Jingxuan Wu, Zhenglin Wan, Xingrui Yu, Yuzhe YANG, Bo An, Ivor Tsang, Yang You
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 orthogonal projection of control and noise to the base velocity (ablation) | OSCAR · Truck concept, CFG 5.0 | BRISQUE: 38.8 -> 21.2 | Improved |
| add redundancy-aware reweighting (ablation) | OSCAR · Truck concept, CFG 5.0 | BRISQUE: 35.0 -> 21.2 | Improved |
| add orthogonal projection (ablation) | OSCAR · Truck concept, CFG 5.0 | FID: 177.8 -> 163.3 | Improved |
| add orthogonal projection (ablation) | OSCAR · Truck concept, CFG 5.0 | Vendi (Inception): 5.06 -> 5.63 | Improved |
| add redundancy-aware reweighting (ablation) | OSCAR · Truck concept, CFG 5.0 | Vendi (Inception): 5.23 -> 5.63 | Improved |
| use intermediate leverage exponent alpha 0.5 instead of 1.0 (ablation) | OSCAR · Truck concept | Brisque: 40.39 -> 20.63 | Improved |
| inject noise over the entire generation process instead of a late window (ablation) | OSCAR · Truck concept | Brisque: 23.13 -> 53.88 | Hurt |
| relax orthogonality strictness lambda from 0.9 to lower values (ablation) | OSCAR · Truck concept | Brisque: 21.56 -> 23.64 (lambda 0.9 -> 0.3) | Hurt |
| apply OSCAR instead of plain flow matching at CFG 7.5 | SD3.5 flow matching · Truck concept, CFG 7.5 | Vendi Score Inception: 3.76 -> 4.08 | Improved |
| apply OSCAR instead of plain flow matching | SD3.5 flow matching · T2I-CompBench spatial color and complex | Vendi (Inception): 3.44 -> 3.96 | Improved |
| use PG latent-space repulsion instead of OSCAR | flow matching diversity guidance · Truck concept, CFG 5.0 | Brisque: 21.72 (OSCAR) vs 40.34 (PG) | Hurt |
| apply OSCAR on FID and CLIP score versus FM-SD3.5 | SD3.5 flow matching · Truck concept, CFG 5.0 | FID: 165.51 -> 164.75 | No clear effect |
Closest papers by meaning
- Don’t Settle at the Mode! Mitigating Diversity Collapse in Pretrained Flow Models via Feature Self-Guidance (ECCV 2026 · similarity 0.91)
- DiverseFlow: Sample-Efficient Diverse Mode Coverage in Flows (CVPR 2025 · similarity 0.90)
- Blockwise Flow Matching: Improving Flow Matching Models For Efficient High-Quality Generation (NeurIPS 2025 · similarity 0.89)
- Safe Few-Step Generation via Velocity Editing (NeurIPS 2026 · similarity 0.89)
- Contextual Flow Matching: Adaptive Step Selection in Flow Models for Efficient Visual Generation (NeurIPS 2026 · similarity 0.89)
- Value Gradient Guidance for Flow Matching Alignment (NeurIPS 2025 · similarity 0.88)
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.
