Research map / Image generation / Fast diffusion sampling
Continuous Adversarial Flow Models
Shanchuan Lin, Ceyuan Yang, Zhijie Lin, Hao Chen, Haoqi Fan
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 |
|---|---|---|---|
| post-train with the CAFM objective | SiT-XL/2 · ImageNet 256, no guidance | FID: 8.26 -> 3.63 | Improved |
| continue training with the FM objective for 10 epochs (control) (ablation) | SiT-XL/2 · ImageNet 256, no guidance | FID: 8.26 -> 8.64 | No clear effect |
| post-train with the CAFM objective | JiT-H/16 · ImageNet 256, no guidance | FID: 7.17 -> 3.57 | Improved |
| continue training with the FM objective for 10 epochs (control) (ablation) | JiT-H/16 · ImageNet 256, no guidance | FID: 7.17 -> 9.30 | Hurt |
| post-train with CAFM, best CFG | SiT-XL/2 · ImageNet 256, guided | FID: 2.06 -> 1.53 | Improved |
| post-train with CAFM, best CFG | JiT-H/16 · ImageNet 256, guided | FID: 1.86 -> 1.80 | Improved |
| post-train with CAFM at high guidance scale CFG 1.6 | SiT-XL/2 · ImageNet 256, CFG 1.6 | FID: 2.25 -> 2.37 | Hurt |
| post-train Z-Image with CAFM instead of FM (no PE, no CFG) | Z-Image · GenEval 512px | overall: 0.33 -> 0.44 | Improved |
| post-train Z-Image with CAFM instead of FM (no CFG) | Z-Image · DPG-Bench 512px | overall: 72.25 -> 77.21 | Improved |
| post-train Z-Image with CAFM instead of FM (with CFG) | Z-Image · DPG-Bench 512px | overall: 83.67 -> 85.21 | Improved |
| use N=16 discriminator updates per generator update instead of N=8 (ablation) | SiT-XL/2 CAFM · ImageNet 256, no guidance | FID: 294.91 -> 3.63 | Improved |
| use N=16 instead of N=32 (ablation) | SiT-XL/2 CAFM · ImageNet 256, no guidance | FID: 3.68 -> 3.63 | No clear effect |
Closest papers by meaning
- Align Your Flow: Scaling Continuous-Time Flow Map Distillation (NeurIPS 2025 · similarity 0.89)
- Self-Corrected Flow Distillation for Consistent One-Step and Few-Step Image Generation (AAAI 2025 · similarity 0.88)
- On the Closed-Form of Flow Matching: Generalization Does Not Arise from Target Stochasticity (NeurIPS 2025 · similarity 0.87)
- Mean Flows for One-step Generative Modeling (NeurIPS 2025 · similarity 0.87)
- Contextual Flow Matching: Adaptive Step Selection in Flow Models for Efficient Visual Generation (NeurIPS 2026 · similarity 0.87)
- A Theory on Flow Matching with Neural Networks (NeurIPS 2026 · 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.
