Research map / Image generation / Autoregressive and latent generation
Trajectory Forcing: Structure-First Generation with Controllable Semantic Trajectories
Merve Kocabas, Gege Gao, Bernhard Schölkopf, Andreas Geiger
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 FD-loss post-training | TF-B/16 · ImageNet 256x256 | FID: 7.93 -> 2.52 | Improved |
| add FD-loss post-training | TF-H/16 · ImageNet 256x256 | FID: 5.97 -> 1.98 | Improved |
| scale model from B/16 to H/16 | TF · ImageNet 256x256 | FID: 7.93 -> 5.97 | Improved |
| structural loss weight lambda=1 instead of 5 or 10 (ablation) | TF-B/16 · ImageNet 256x256 | FID | Improved |
| use Muon instead of AdamW (ablation) | TF-B/16 · ImageNet 256x256 | L-NFE FID during training | Improved |
| edit at coarser hierarchy level (ablation) | TF · Masked SSIM on unedited pixels (feature edits) | masked SSIM: 0.92 (subparts) -> 0.77 (object/background) | Hurt |
| use 4-step hierarchical one-step flow instead of many-step baselines | TF · ImageNet 256x256 | FID at 80 epochs: TF-B/16 7.93 vs iMF-L/2 1.86 | Mixed |
| use shared RAE decoder across levels without per-level finetuning | TF · intermediate state decoding | decodability | Improved |
Closest papers by meaning
- Role Bias in Diffusion Models: Diagnosing and Mitigating through Intermediate Decomposition (NeurIPS 2025 · similarity 0.87)
- Contextual Flow Matching: Adaptive Step Selection in Flow Models for Efficient Visual Generation (NeurIPS 2026 · similarity 0.87)
- Tracing the Roots: Leveraging Temporal Dynamics in Diffusion Trajectories for Origin Attribution (NeurIPS 2025 · similarity 0.87)
- FoMo: Forking Moment in Generative Trajectory as a Perceptual Distance (NeurIPS 2026 · similarity 0.87)
- Drag Your Noise: Interactive Point-based Editing via Diffusion Semantic Propagation (CVPR 2024 · similarity 0.87)
- Deep Compositional Phase Diffusion for Long Motion Sequence Generation (NeurIPS 2025 · similarity 0.86)
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.
