Research map / Image generation / Fast diffusion sampling
HierarchicalPrune: Position-Aware Compression for Large-Scale Diffusion Models
Young D. Kwon, Rui Li, Sijia Li, Da Li, Sourav Bhattacharya, Stylianos I. Venieris
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 positional weight preservation (PWP) to HPP, 20% pruning (ablation) | HierarchicalPrune on SD3.5 Large Turbo · GenEval | GenEval score: 0.03 -> 0.71 | Improved |
| add PWP to HPP, 20% pruning (ablation) | HierarchicalPrune on SD3.5 Large Turbo · HPSv2 | HPSv2 score: 11.08 -> 28.97 | Improved |
| add PWP to HPP, 30% pruning (ablation) | HierarchicalPrune on SD3.5 Large Turbo · GenEval | GenEval score: 0.0 -> 0.46 | Improved |
| add sensitivity-guided distillation (SGDistill) to HPP+PWP, 30% pruning (ablation) | HierarchicalPrune on SD3.5 Large Turbo · GenEval | GenEval score: 0.46 -> 0.64 | Improved |
| add SGDistill to HPP+PWP, 30% pruning (ablation) | HierarchicalPrune on SD3.5 Large Turbo · HPSv2 | HPSv2 score: 21.74 -> 27.29 | Improved |
| add W4A16 quantisation at 20% pruning (ablation) | HierarchicalPrune on SD3.5 Large Turbo · GenEval | GenEval score: 0.71 -> 0.69 | Hurt |
| add W4A16 quantisation at 30% pruning (ablation) | HierarchicalPrune on SD3.5 Large Turbo · HPSv2 | HPSv2 score: 27.29 -> 26.29 | Hurt |
| use HierarchicalPrune (HPP+PWP+Q) instead of KOALA | SD3.5 Large Turbo · HPSv2 | HPSv2 score: 19.99 -> 28.15 | Improved |
| use HierarchicalPrune instead of BK-SDM on FLUX.1 Schnell | FLUX.1 Schnell · GenEval | GenEval score: 0.45 -> 0.64 | Improved |
| use HierarchicalPrune instead of KOALA/BK-SDM on SD3.5 for latency | SD3.5 Large Turbo · Per-step latency (A6000) | latency (ms): 642 ms -> 593 ms | Improved |
| use HierarchicalPrune instead of KOALA/BK-SDM on FLUX.1 for latency | FLUX.1 Schnell · Per-step latency (A6000) | latency (ms): 432 ms -> 469 ms | Hurt |
Closest papers by meaning
- Würstchen: An Efficient Architecture for Large-Scale Text-to-Image Diffusion Models (ICLR 2024 · similarity 0.91)
- OSCAR: One-Step Diffusion Codec Across Multiple Bit-rates (NeurIPS 2025 · similarity 0.89)
- Exploiting Diffusion Prior for Generalizable Dense Prediction (CVPR 2024 · similarity 0.89)
- SANA 1.5: Efficient Scaling of Training-Time and Inference-Time Compute in Linear Diffusion Transformer (ICML 2025 · similarity 0.89)
- On improved Conditioning Mechanisms and Pre-training Strategies for Diffusion Models (NeurIPS 2024 · similarity 0.88)
- Faster Diffusion: Rethinking the Role of the Encoder for Diffusion Model Inference (NeurIPS 2024 · 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.
