Research map / Image generation / Fast diffusion sampling
AegisFlow: Training-free Non-myopic Path-safe Guided Flow Matching
Kunpeng Liu, Boshi Zhang, Keyou You
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes AegisFlow, a training-free, non-myopic guidance method for pretrained flow matching and diffusion models that steers generative trajectories to satisfy path-safety constraints such as identity preservation and obstacle avoidance.
Problem
Steering pretrained diffusion and flow matching generative trajectories to satisfy strict constraints (e.g., identity preservation in image editing, obstacle avoidance in robotics) is hard; existing training-free guidance methods enforce path-safety but, per the abstract, rely on (truncated) myopic
Approach
Non-myopic, path-safe guidance applied training-free to flow matching sampling so that constraints hold along the whole generative trajectory rather than only at each local step. Details beyond the abstract are not available.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace myopic min-norm controller with non-myopic barrier-cost HJB control | OC-Flow on CelebA-HQ · CelebA-HQ text-guided editing | CLIP: 0.304 -> 0.309 | Improved |
| replace myopic min-norm controller with non-myopic barrier-cost HJB control | OC-Flow on CelebA-HQ · CelebA-HQ text-guided editing | ID: 0.920 -> 0.901 | Hurt |
| add MNC safeguard to OC-Flow | OC-Flow · CelebA-HQ text-guided editing | ID: 0.732 -> 0.920 | Improved |
| AegisFlow vs SafeFlowMatcher (MNC) on maze planning | flow matching planner · Maze2D maze_large_v1 | average curvature: 51.97 -> 44.72 | Improved |
| AegisFlow vs SafeFlowMatcher on maze planning | flow matching planner · Maze2D maze_large_v1 | BS2: 0.041 -> 0.071 | Improved |
| AegisFlow vs SafeFlowMatcher on maze planning | flow matching planner · Maze2D maze_large_v1 | D4RL normalized score: 1.632 -> 1.635 | No clear effect |
| AegisFlow vs DPCC-C on robot avoidance | flow matching policy · D3IL Avoiding | timesteps to goal: 69.0 -> 62.6 | Improved |
| AegisFlow vs DPCC-C on robot avoidance | flow matching policy · D3IL Avoiding | safe success rate: 0.98 -> 0.98 | No clear effect |
| lower control cost weight lambda from 1.0 to 0.4 (ablation) | AegisFlow · CelebA-HQ text-guided editing | CLIP: 0.299 -> 0.316 | Improved |
| lower control cost weight lambda from 1.0 to 0.4 (ablation) | AegisFlow · CelebA-HQ text-guided editing | LPIPS: 0.152 -> 0.224 | Hurt |
| lower control cost weight lambda from 1.0 to 0.4 (ablation) | AegisFlow · CelebA-HQ text-guided editing | ID: 0.903 -> 0.903 | No clear effect |
| replace MNC with HJB-based control | OC-Flow · CelebA-HQ computation time | seconds: 518.26 -> 433.81 | Improved |
Closest papers by meaning
- On the Guidance of Flow Matching (ICML 2025 · similarity 0.89)
- Aligning Flow Map Policies with Optimal $Q$-Guidance (NeurIPS 2026 · similarity 0.87)
- Generative Control as Optimization: Time Unconditional Flow Matching for Adaptive and Robust Robotic Control (NeurIPS 2026 · similarity 0.87)
- FlowPolicy: Enabling Fast and Robust 3D Flow-Based Policy via Consistency Flow Matching for Robot Manipulation (AAAI 2025 · similarity 0.87)
- Training Free Guided Flow-Matching with Optimal Control (ICLR 2025 · similarity 0.87)
- Optimal Flow Matching: Learning Straight Trajectories in Just One Step (NeurIPS 2024 · 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.
