Research map / Image generation / Fast diffusion sampling
Q-Sched: Pushing the Boundaries of Few-Step Diffusion Models with Quantization-Aware Scheduling
Natalia Frumkin, Diana Marculescu
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 Q-Sched scheduler instead of the original schedule under W4A8 | 2-step PCM (SD v1-5) · COCO-30k | FID: 24.17 -> 22.24 | Improved |
| apply Q-Sched scheduler instead of the original schedule under W4A8 | 2-step LCM (SD v1-5) · COCO-30k | FID: 40.93 -> 32.50 | Improved |
| Q-Sched instead of PTQD noise schedule under W4A8 | 8-step PCM (SD v1-5) · COCO-30k | FID: 15.85 -> 16.83 | Hurt |
| Q-Sched instead of PTQD on a large SDXL backbone | 2-step PCM (SDXL) W4A8 · COCO-30k | FID: 161.96 -> 18.89 | Improved |
| Q-Sched instead of naive quantization at W4A8 | SDXL-Turbo · MJHQ | FID: 25.75 -> 21.41 | Improved |
| Q-Sched instead of naive quantization at W8A8 | SDXL-Turbo · MJHQ | FID: 25.49 -> 26.34 | Hurt |
| Q-Sched instead of PTQD with added stochasticity eta=0.9 (ablation) | 2-step PCM (SD v1-5) W4A8 · COCO-30k | FID: 26.71 -> 17.87 | Improved |
| Q-Sched instead of TCD at eta=0.3 (ablation) | 2-step PCM (SD v1-5) W4A8 · COCO-30k | FID: 23.44 -> 23.44 | No clear effect |
| jointly optimize both model-side c_eps and sample-side c_x coefficients rather than one (ablation) | Q-Sched on PCM (SDXL) · 1024 generated images | PickScore, HPSv2, JAQ loss | Improved |
| use large JAQ trade-off k=5 instead of k=2 (ablation) | Q-Sched · qualitative | image fidelity | Hurt |
| optimize Q-Sched with CLIPScore alone instead of JAQ loss (ablation) | Q-Sched · qualitative | image quality | Hurt |
| optimize with Brisque instead of JAQ loss (ablation) | Q-Sched · qualitative | image detail | Hurt |
Closest papers by meaning
- Q-DiT: Accurate Post-Training Quantization for Diffusion Transformers (CVPR 2025 · similarity 0.91)
- EfficientDM: Efficient Quantization-Aware Fine-Tuning of Low-Bit Diffusion Models (ICLR 2024 · similarity 0.91)
- TCAQ-DM: Timestep-Channel Adaptive Quantization for Diffusion Models (AAAI 2025 · similarity 0.90)
- PTQ4DiT: Post-training Quantization for Diffusion Transformers (NeurIPS 2024 · similarity 0.90)
- Towards Accurate Post-training Quantization for Diffusion Models (CVPR 2024 · similarity 0.89)
- ViDiT-Q: Efficient and Accurate Quantization of Diffusion Transformers for Image and Video Generation (ICLR 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.
