Research map / Image generation / Fast diffusion sampling
Ctrl-Z Sampling: Scaling Diffusion Sampling with Controlled Random Zigzag Explorations
Shunqi Mao, Wei Guo, Chaoyi Zhang, Jieting Long, Ke Xie, Weidong Cai
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 |
|---|---|---|---|
| trigger exploration on reward plateau instead of at every step (ablation) | Ctrl-Z Sampling · Pick-a-Pic (SD 2.1) | ImageReward / NFEs: IR 1.317 -> 1.138; NFEs 16.72 -> 7.72 | Mixed |
| trigger exploration on reward plateau instead of randomly with p=0.5 (ablation) | Ctrl-Z Sampling · Pick-a-Pic (SD 2.1) | ImageReward: 1.062 -> 1.138 | Improved |
| increase exploration probability from p=0 to p=1 (ablation) | Ctrl-Z Sampling · Pick-a-Pic (SD 2.1) | ImageReward / NFEs: IR 0.194 -> 1.317; NFEs 1.00 -> 16.72 | Improved |
| widen exploration window lambda from 0/50 to 40/50 (ablation) | Ctrl-Z Sampling · Pick-a-Pic (SD 2.1) | ImageReward: 0.194 -> 1.138 | Improved |
| widen exploration window lambda from 40/50 to 50/50 (ablation) | Ctrl-Z Sampling · Pick-a-Pic (SD 2.1) | HPS v2 / IR / NFEs: HPS v2 27.34 -> 27.29; IR 1.138 -> 1.286; NFEs 7.72 -> 10.19 | Mixed |
| use Aesthetic Score instead of ImageReward as reward model (ablation) | Ctrl-Z Sampling · Pick-a-Pic (SD 2.1) | AES / IR: AES 5.705 -> 6.147; IR 1.138 -> 0.306 | Mixed |
| use PickScore instead of ImageReward as reward model (ablation) | Ctrl-Z Sampling · Pick-a-Pic (SD 2.1) | PickScore / IR: PickScore 21.02 -> 21.69; IR 1.138 -> 0.589 | Mixed |
| use Ctrl-Z (default) instead of SOP-4 | SOP-4 · Pick-a-Pic (SD 2.1) | HPS v2 / NFEs: HPS v2 27.23 -> 27.34; NFEs 9.00 -> 7.72 | Improved |
| use Ctrl-Z (low NFE) instead of SOP-1 | SOP-1 · Pick-a-Pic (SD 2.1) | ImageReward / NFEs: IR 0.735 -> 0.720; NFEs 3.00 -> 2.77 | No clear effect |
| use Ctrl-Z (default) instead of DDIM | DDIM · DrawBench (Hunyuan-DiT) | ImageReward: 0.827 -> 1.319 | Improved |
| use Ctrl-Z (default) instead of DDIM | DDIM · T2I-CompBench Color (SD 2.1) | attribute score: 46.27 -> 61.26 | Improved |
| use Ctrl-Z (default) instead of SOP-4 | SOP-4 · T2I-CompBench Numeracy (Hunyuan-DiT) | numeracy score: 57.58 vs 56.69 | No clear effect |
Closest papers by meaning
- Test-Time Scaling of Diffusion Models via Noise Trajectory Search (NeurIPS 2025 · similarity 0.92)
- CCS: Controllable and Constrained Sampling with Diffusion Models via Initial Noise Perturbation (NeurIPS 2025 · similarity 0.90)
- Provable Posterior Sampling with Denoising Oracles via Tilted Transport (NeurIPS 2024 · similarity 0.90)
- Divide-and-Conquer Posterior Sampling for Denoising Diffusion priors (NeurIPS 2024 · similarity 0.90)
- Efficient Denoising Diffusion via Probabilistic Masking (ICML 2024 · similarity 0.90)
- EVODiff: Entropy-aware Variance Optimized Diffusion Inference (NeurIPS 2025 · similarity 0.89)
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.
