Research map / Image generation / Fast diffusion sampling
Leveraging Early-Stage Robustness in Diffusion Models for Efficient and High-Quality Image Synthesis
Yulhwa Kim, Dongwon Jo, Hyesung Jeon, Taesu Kim, Daehyun Ahn, Hyungjun Kim, jae-joon kim
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 |
|---|---|---|---|
| apply 4-bit activation quantization only in early steps 1.0T:0.8T instead of late steps 0.2T:0.0T (ablation) | LDM-8 W4 activation quantization · LSUN-Churches 256x256 unconditional (5k images) | FID: 34.30 -> 6.10 | Improved |
| use step-wise activation bits (4b early to 8b late, effective 6) instead of uniform 6-bit (ablation) | RAQ LDM-8 · LSUN-Churches 256x256 unconditional | FID: 5.31 -> 4.64 | Improved |
| use step-wise activation bits (effective 6) instead of uniform 6-bit (ablation) | RAQ LDM-4 · LSUN-Bedrooms 256x256 unconditional | FID: 5.58 -> 3.99 | Improved |
| replace Q-diffusion W4/A8 with RAQ at effective 6-bit activations | LDM-8 · LSUN-Churches 256x256 unconditional | TBOPs: 148.6 -> 108.8 | Improved |
| replace Q-diffusion W4/A8 with RAQ at effective 6-bit activations | LDM-4 · LSUN-Bedrooms 256x256 unconditional | TBOPs: 681.2 -> 504.8 | Improved |
| replace Q-diffusion W4/A8 with RAQ at effective 6-bit activations | LDM-8 · LSUN-Churches 256x256 unconditional | FID: 4.45 -> 4.64 | No clear effect |
| replace Q-diffusion W4/A8 with RAQ at effective 6-bit activations | LDM-4 · LSUN-Bedrooms 256x256 unconditional | FID: 4.17 -> 3.99 | Improved |
| lower effective activation bits further to 5.60 / 5.35 (relaxed FID) (ablation) | RAQ LDM-8 · LSUN-Churches 256x256 unconditional | FID: 4.64 -> 5.12 -> 5.51 | Hurt |
| lower effective activation bits further to 5.50 / 5.45 (ablation) | RAQ LDM-4 · LSUN-Bedrooms 256x256 unconditional | FID: 3.99 -> 4.94 -> 5.32 | Hurt |
| lower effective activation bits further to 5.60 / 5.35 (ablation) | RAQ LDM-8 · LSUN-Churches 256x256 unconditional | TBOPs: 108.8 -> 101.3 -> 96.4 | Improved |
| reduce weight bits down to 4 (weight-only quantization) (ablation) | LDM-8 BRECQ · LSUN-Churches 256x256 unconditional | FID | No clear effect |
| reduce activation bits below 8 (ablation) | LDM-8 · LSUN-Churches 256x256 unconditional | FID | Hurt |
Closest papers by meaning
- EfficientDM: Efficient Quantization-Aware Fine-Tuning of Low-Bit Diffusion Models (ICLR 2024 · similarity 0.91)
- Optimizing Quantized Diffusion Models via Distillation with Cross-Timestep Error Correction (AAAI 2025 · similarity 0.90)
- D^2-DPM: Dual Denoising for Quantized Diffusion Probabilistic Models (AAAI 2025 · similarity 0.90)
- A Simple Early Exiting Framework for Accelerated Sampling in Diffusion Models (ICML 2024 · similarity 0.90)
- Relay Diffusion: Unifying diffusion process across resolutions for image synthesis (ICLR 2024 · similarity 0.90)
- Simple Distillation for One-Step Diffusion Models (NeurIPS 2025 · similarity 0.90)
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.
