Research map / Image generation / Fast diffusion sampling
From Predictors to Samplers via the Training Trajectory
Soumya Ram, Akhila Ram
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 |
|---|---|---|---|
| anneal across training checkpoints instead of final checkpoint only (ablation) | GWG sampling · Boolean polynomial target | success rate: 0.04 -> 0.5200 | Improved |
| anneal across training checkpoints instead of final checkpoint only (ablation) | GWG sampling · Indicator d=10 | success probability: 0.12 -> 0.99 | Improved |
| increase indicator size d from 3 to 10 (ablation) | GWG on final checkpoint · Indicator function | success probability: 0.47 -> 0.12 | Hurt |
| increase indicator size d from 3 to 10 (ablation) | trajectory annealing · Indicator function | success probability: 0.98 -> 0.99 | No clear effect |
| anneal across training checkpoints with adversarial linear term (ablation) | GWG sampling · Non-convex binary landscape | success rate: 0.080 -> 1.00 | Improved |
| anneal across training checkpoints with three length-5 indicators (ablation) | GWG sampling · 3 length-5 indicators | success rate: 0.0250 -> 1.00 | Improved |
| anneal across training checkpoints (1K steps) (ablation) | Temp-GWG · Binary MNIST EBM | FID: 29.61 -> 11.73 | Improved |
| anneal across training checkpoints (10K steps) (ablation) | Temp-GWG · Binary MNIST EBM | FID: 21.12 -> 5.49 | Improved |
| use 5 checkpoints instead of baseline final-checkpoint annealing (ablation) | trajectory annealing · Binary MNIST EBM, 1K steps | FID: 29.61 -> 16.10 | Improved |
| increase number of checkpoints from 5 to 500 (ablation) | trajectory annealing · Binary MNIST EBM, 1K steps | FID: 16.10 -> 11.73 | Improved |
| anneal across training checkpoints (ablation) | GWG sampling · DNA design (MAX TF) | median fitness: 2.72 -> 10.04 | Improved |
| anneal across training checkpoints under Hamming-7 constraint (ablation) | GWG sampling · Constrained DNA design | motif (%): 31.3 -> 63.0 | Improved |
Closest papers by meaning
- Scalable Discrete Diffusion Samplers: Combinatorial Optimization and Statistical Physics (ICLR 2025 · similarity 0.89)
- Consistent Diffusion Meets Tweedie: Training Exact Ambient Diffusion Models with Noisy Data (ICML 2024 · similarity 0.88)
- Diffusion Generative Flow Samplers: Improving learning signals through partial trajectory optimization (ICLR 2024 · similarity 0.88)
- Tracing the Roots: Leveraging Temporal Dynamics in Diffusion Trajectories for Origin Attribution (NeurIPS 2025 · similarity 0.88)
- Your Diffusion Model is Secretly a Noise Classifier and Benefits from Contrastive Training (NeurIPS 2024 · similarity 0.88)
- When Diffusion Models Memorize: Inductive Biases in Probability Flow of Minimum-Norm Shallow Neural Nets (ICML 2025 · similarity 0.88)
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.
