Research map / Image generation / Fast diffusion sampling
FlowSteer: Guiding Few-Step Image Synthesis with Authentic Trajectories
Lei Ke, Hubery Yin, Gongye Liu, Zhengyao Lv, Jingcai Guo, Chen Li, Wenhan Luo, Yujiu Yang, Jing Lyu
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 |
|---|---|---|---|
| Use authentic teacher trajectory start point instead of linear interpolation | SD3 teacher · COCO 1k prompts | HPSv2: 28.63 -> 29.43 | Improved |
| Add Online Trajectory Alignment (ablation) | PeRFlow baseline, SD3-Medium · SD3M 4-step ablation | HPSv2: 26.36 -> 26.79 | Improved |
| Add adversarial distillation on ODE trajectory (ablation) | PeRFlow baseline, SD3-Medium · SD3M 4-step ablation | HPSv2: 26.36 -> 27.73 | Improved |
| Add adversarial distillation on ODE trajectory (ablation) | PeRFlow baseline, SD3-Medium · SD3M 4-step ablation | PickScore: 22.19 -> 22.18 | No clear effect |
| Add improved scheduler (ablation) | PeRFlow baseline, SD3-Medium · SD3M 4-step ablation | PickScore: 22.19 -> 22.50 | Improved |
| Combine OTA, adversarial distillation and scheduler (full method) (ablation) | PeRFlow baseline, SD3-Medium · SD3M 4-step ablation | HPSv2: 26.36 -> 28.60 | Improved |
| Use improved scheduler with linear spacing including terminal sigma | SD3M original scheduler, shift 3, 4 steps · SD3M few-step inference | HPSv2: 15.91 -> 19.34 | Improved |
| Use improved scheduler at 32 steps | SD3M, shift 3 · SD3M inference | HPSv2: 29.50 -> 29.51 | No clear effect |
| Use 12-block discriminator backbone instead of 6, 18 or 24 blocks (ablation) | FlowSteer discriminator · SD3M 4-step | HPSv2: 28.25 (24 blocks) -> 28.60 (12 blocks) | Improved |
| Use one discriminator head on last block instead of per-block heads (ablation) | FlowSteer discriminator · SD3M 4-step | HPSv2: 27.44 -> 28.60 | Improved |
| Fully fine-tune discriminator backbone instead of frozen or LoRA (ablation) | FlowSteer discriminator · SD3M 4-step | HPSv2: 27.45 (frozen) / 28.07 (LoRA) -> 28.60 | Improved |
| Use hinge loss instead of LSGAN or WGAN (ablation) | FlowSteer adversarial loss · SD3M 4-step | HPSv2: 27.64 (LSGAN), 27.28 (WGAN) -> 28.60 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- PeRFlow — PeRFlow: Piecewise Rectified Flow as Universal Plug-and-Play Accelerator (NeurIPS 2024) · better on 6 of 6 shared results
- Flash Diffusion — Flash Diffusion: Accelerating Any Conditional Diffusion Model for Few Steps Image Generation (AAAI 2025) · better on 3 of 3 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
- Self-Corrected Flow Distillation for Consistent One-Step and Few-Step Image Generation (AAAI 2025 · similarity 0.91)
- Align Your Flow: Scaling Continuous-Time Flow Map Distillation (NeurIPS 2025 · similarity 0.90)
- Beyond Trajectory Matching: Reflow with Marginal Distribution Alignment (NeurIPS 2026 · similarity 0.90)
- FlowTurbo: Towards Real-time Flow-Based Image Generation with Velocity Refiner (NeurIPS 2024 · similarity 0.89)
- Contextual Flow Matching: Adaptive Step Selection in Flow Models for Efficient Visual Generation (NeurIPS 2026 · similarity 0.89)
- Bellman Optimal Stepsize Straightening of Flow-Matching Models (ICLR 2024 · 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.
