Research map / Image generation / Fast diffusion sampling
Scaling Group Inference for Diverse and High-Quality Generation
Gaurav Parmar, Or Patashnik, Daniil Ostashev, Kuan-Chieh Jackson Wang, Kfir Aberman, Srinivasa Narasimhan, Jun-Yan Zhu
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 progressive pruning with intermediate predictions (ablation) | group inference · text-to-image group score | runtime: up to 73% less runtime | Improved |
| progressive pruning vs full denoising of all candidates (ablation) | group inference · FLUX.1 model variants | speedup: 49%, 73%, and 55% faster | Improved |
| prune candidates with ratio 0.5 instead of 0.1 or 0.25 (ablation) | group inference · FLUX.1 Dev | combined score | Improved |
| aggressive pruning rho=0.1 (ablation) | group inference · FLUX.1 Schnell | combined score | No clear effect |
| add efficient decoder for intermediate predictions (ablation) | group inference · FLUX.1 Dev and Schnell | runtime | Improved |
| increase number of denoising steps | independent sampling · group objective | combined score | No clear effect |
| random seed search with CLIP verifier | Inference Diffusion Scaling · group objective | combined score | No clear effect |
| increase initial samples M from 4 to 128 | group inference · 3 base models | quality and diversity | Improved |
| use low CFG (CFG=1) to gain diversity | FLUX.1 Dev · text-to-image | quality: 88.3% vs 11.70% diversity pref; 85.6% vs 14.4% quality pref for ours | Hurt |
| add particle guidance | FLUX.1 Dev · text-to-image | quality: 79.4% vs 20.6% quality pref for ours | Hurt |
| interval guidance vs ours | FLUX.1 Dev · text-to-image | user preference: 53.4% vs 46.6% diversity; 58.4% vs 41.6% quality | Improved |
| replace binary term with color dissimilarity | group inference · giant neon rose prompt | diversity | Improved |
Closest papers by meaning
- Improving the Scaling Laws of Synthetic Data with Deliberate Practice (ICML 2025 · similarity 0.85)
- Project-Probe-Aggregate: Efficient Fine-Tuning for Group Robustness (CVPR 2025 · similarity 0.85)
- Flexible Group Count Enables Hassle-Free Structured Pruning (CVPR 2025 · similarity 0.85)
- Multi-Group Proportional Representations for Text-to-Image Models (CVPR 2025 · similarity 0.85)
- Scaling Inference Time Compute for Diffusion Models (CVPR 2025 · similarity 0.85)
- PermuQuant: Lowering Per-Group Quantization Error by Reordering Channels for Diffusion Models (NeurIPS 2026 · similarity 0.85)
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.
