Research map / Image generation / Fast diffusion sampling
Inference-Time Scaling of Discrete Diffusion Models via Importance Weighting and Optimal Proposal Design
Zijing Ou, Chinmay Pani, Yingzhen Li
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 |
|---|---|---|---|
| run SMC with 8 particles on top of the amortised proposal | Prop_amot · Toxic text generation (MDLM) | Toxic rate: 63.7% -> 100.0% | Improved |
| run SMC with 8 particles on top of the gradient proposal | Prop_grad · Toxic text generation (MDLM) | Toxic rate: 58.0% -> 95.0% | Improved |
| use the approximated optimal proposal instead of the pretrained model as SMC proposal (grad) | SMC (N=8) · Toxic text generation (MDLM) | Toxic rate: 26.7% -> 95.0% | Improved |
| use SMC instead of best-of-N with 8 samples | BoN · Toxic text generation (MDLM) | Toxic rate: 6.3% -> 26.7% | Improved |
| use the learned amortised proposal instead of the grad proposal in SMC (ablation) | SMC (N=8) · Toxic text generation (MDLM) | GPT2-XL perplexity: 132.1 -> 147.6 | Hurt |
| use the learned amortised proposal instead of the grad proposal in SMC (ablation) | SMC (N=8) · Toxic text generation (MDLM) | Dist-1/2/3: 57/92/96 -> 44/81/91 | Hurt |
| train the amortised proposal with log-variance minimisation instead of gradient approximation (ablation) | Prop (N=1) · Toxic text generation (MDLM) | Toxic (Holdout) rate: 58.3% -> 75.7% | Improved |
| increase particles from 1 to 16 at 8 steps (ablation) | SMC on MaskGit · ImageNet256 (CFG=1.25, 8 steps) | FID: 24.64 -> 14.56 | Improved |
| increase particles from 1 to 16 at 32 steps (ablation) | SMC on MaskGit · ImageNet256 (CFG=1.25, 32 steps) | FID: 12.02 -> 8.76 | Improved |
| increase particles from 1 to 16 at 8 steps (ablation) | SMC on MaskGit · ImageNet256 (CFG=1.25, 8 steps) | Inception Score: 62.8 -> 107.4 | Improved |
| increase the number of particles in DNA sequence design (ablation) | SMC · DNA sequence design | Pred-Activity and ATAC-Acc | Improved |
| increase the number of particles in text-to-image generation (ablation) | SMC (Meissonic) · Meissonic reward alignment | reward | Improved |
Closest papers by meaning
- Scalable Discrete Diffusion Samplers: Combinatorial Optimization and Statistical Physics (ICLR 2025 · similarity 0.90)
- Fast Solvers for Discrete Diffusion Models: Theory and Applications of High-Order Algorithms (NeurIPS 2025 · similarity 0.90)
- Scaling Inference Time Compute for Diffusion Models (CVPR 2025 · similarity 0.89)
- Remasking Discrete Diffusion Models with Inference-Time Scaling (NeurIPS 2025 · similarity 0.89)
- Simple Guidance Mechanisms for Discrete Diffusion Models (ICLR 2025 · similarity 0.89)
- A Mean-Field Framework for Inference-Time Distributional Control of Diffusion Models (NeurIPS 2026 · 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.
