Research map / Image generation / Fast diffusion sampling
Conflict-Aware Additive Guidance for Flow Models under Compositional Rewards
Xuehui Yu, Fucheng Cai, WANG MEIYI, Xiaopeng Fan, Harold Soh
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 conflict-aware learned guidance (gcar) | gcov-G approximate guidance · Synthetic 2D MoG [1,0] | posterior coverage (%): 71.70 -> 93.80 | Improved |
| add PCGrad gradient projection | gcov-G · Synthetic 2D MoG [1,0] | posterior coverage (%): 71.70 -> 75.20 | Improved |
| add gcar | gcov-G · Maze2D (4 compositional settings) | success rate (%): 16 -> 79 (static obstacles) | Improved |
| add PCGrad | gcov-G · Maze2D | success rate (%): 16 -> 27 (static obstacles) | Mixed |
| add gcar on top of MPPI | MPPI · Maze2D static obstacles | success rate (%): 41 -> 98 | Improved |
| add gcar | gcov-G · ManiSkill2 StackCube hybrid composition | success rate (%): 9 -> 61 | Improved |
| add PCGrad | gcov-G · ManiSkill2 StackCube | success rate (%): 12 -> 0 (static obstacles) | Hurt |
| add gcar | gcov-G · CelebA-HQ composed prompts | ID similarity: 0.543 -> 0.681 | Improved |
| add gcar | gcov-G · CelebA-HQ composed prompts | LPIPS: 0.217 -> 0.226 | Hurt |
| use gcar instead of exact GM guidance | GM · Synthetic 2D MoG | training data usage: 10240 -> 549.55 (x10^3) | Improved |
| use gcar instead of sample-based GLASS-FKS | GLASS-FKS (K=16) · Synthetic 2D MoG [1,0] | inference time (ms/sample): 296 -> 4.20 | Improved |
| lower early-stopping threshold eps from 0.10 to 0.00 (ablation) | gcar · Synthetic 2D MoG avg composed | posterior coverage (%): 88.80 -> 95.10 | Improved |
Closest papers by meaning
- Greed is Good: A Unifying Perspective on Guided Generation (NeurIPS 2025 · similarity 0.89)
- Temporal Alignment Guidance: On-manifold Sampling in Diffusion Models (NeurIPS 2026 · similarity 0.88)
- On the Guidance of Flow Matching (ICML 2025 · similarity 0.88)
- Token Perturbation Guidance for Diffusion Models (NeurIPS 2025 · similarity 0.87)
- Local Manifold Approximation and Projection for Manifold-Aware Diffusion Planning (ICML 2025 · similarity 0.87)
- Tortoise and Hare Guidance: Accelerating Diffusion Model Inference with Multirate Integration (NeurIPS 2025 · 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.
