Research map / Image generation / Fast diffusion sampling
"BK-SDM: A Lightweight, Fast, and Cheap Version of Stable Diffusion"
Bo-Kyeong Kim, Hyoung-Kyu Song, Thibault Castells, Shinkook Choi
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 |
|---|---|---|---|
| remove blocks from U-Net and retrain with distillation (BK-SDM-Base) | SDM-v1.4 · GPU latency, 512x512 | whole-pipeline latency: 1.41s -> 0.96s for Base | Improved |
| add distillation (output + feature KD) during retraining (ablation) | BK-SDM-Base · MS-COCO 30K | FID: 23.57 -> 15.76 | Improved |
| add output-level KD on top of denoising task loss (ablation) | BK-SDM-v2-Small (teacher init) · MS-COCO 30K | FID: 16.71 -> 14.27 | Improved |
| add feature-level KD on top of output KD (ablation) | BK-SDM-v2-Small (teacher init) · MS-COCO 30K | FID: 14.27 -> 16.61 FID; CLIP 0.2777 -> 0.2901 | Mixed |
| initialize from teacher weights instead of random (ablation) | BK-SDM-v2-Small · MS-COCO 30K | FID: 24.47 -> 16.61 | Improved |
| use magnitude pruning criterion instead of hand-designed block removal (ablation) | BK-SDM-Small · MS-COCO 30K | FID: >150 | Hurt |
| use Taylor pruning criterion instead of hand-designed removal (ablation) | BK-SDM-Small · MS-COCO 30K | U-Net MACs: 217.7G -> 305.4G MACs; FID 16.83 -> 16.73 | Hurt |
| use CLIP-score sensitivity criterion instead of hand-designed removal (ablation) | BK-SDM-Small · MS-COCO 30K | FID: 16.83 -> 21.86 | Hurt |
| increase batch size from 64 to 256 (ablation) | BK-SDM-Base · MS-COCO 30K | FID: FID 14.61 -> 15.76; IS 31.44 -> 33.79 | Mixed |
| remove the entire mid-stage of the U-Net without retraining (ablation) | SDM-v1.4 · MS-COCO 256x256 30K | FID: 13.05 -> 15.60; U-Net params -11.3% | Mixed |
| remove no KD from personalization finetuning backbone (add KD retraining) (ablation) | BK-SDM-Base DreamBooth · DreamBooth | CLIP-I: 0.669 -> 0.708 | Improved |
| use BK-SDM backbone for DreamBooth finetuning instead of SDM-v1.4 | SDM-v1.4 · DreamBooth | DINO: 0.728 -> 0.723 (Base) | No clear effect |
Later papers that compared with this method
5 later accepted papers report BK-SDM Base as a baseline in their results tables; 5 of them report a better number on most of the results they share with it.
- 2ndMatch: Finetuning Pruned Diffusion Models via Second-Order Jacobian Matching (CVPR 2026 · better on 3 of 4 shared results)
- Random Conditioning for Diffusion Model Compression with Distillation (CVPR 2025 · better on 9 of 9 shared results)
- Random Conditioning with Distillation for Data-Efficient Diffusion Model Compression (CVPR 2025 · better on 3 of 3 shared results)
- ObCLIP: Oblivious CLoud-Device Hybrid Image Generation with Privacy Preservation (NeurIPS 2025 · better on 1 of 1 shared result)
- DeepCache: Accelerating Diffusion Models for Free (CVPR 2024 · better on 2 of 2 shared results)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- KOALA: Empirical Lessons Toward Memory-Efficient and Fast Diffusion Models for Text-to-Image Synthesis (NeurIPS 2024 · similarity 0.90)
- Faster Diffusion: Rethinking the Role of the Encoder for Diffusion Model Inference (NeurIPS 2024 · similarity 0.89)
- Not All Prompts Are Made Equal: Prompt-based Pruning of Text-to-Image Diffusion Models (ICLR 2025 · similarity 0.88)
- On the Scalability of Diffusion-based Text-to-Image Generation (CVPR 2024 · similarity 0.88)
- InstaFlow: One Step is Enough for High-Quality Diffusion-Based Text-to-Image Generation (ICLR 2024 · similarity 0.88)
- CommonCanvas: Open Diffusion Models Trained on Creative-Commons Images (CVPR 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.
