Research map / Image generation / Fast diffusion sampling
LogCD: Local-to-global Consistency Distillation for Few-step Image Generation
Qingsong Xie, Zhenyi Liao, Chen Chen, Zhijie Deng, Haonan Lu
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 |
|---|---|---|---|
| replace MSE with latent LPIPS loss in local CD (ablation) | LogCD (SDXL) · MSCOCO-2017 5K | ImageReward: 0.52 -> 0.62 | Improved |
| add global consistency distillation after local CD (ablation) | LogCD (SDXL) · MSCOCO-2017 5K | ImageReward: 0.62 -> 0.76 | Improved |
| add GAN loss to global CD (ablation) | LogCD (SDXL) · MSCOCO-2017 5K | ImageReward: 0.76 -> 0.82 | Improved |
| add DMD loss to global CD (ablation) | LogCD (SDXL) · MSCOCO-2017 5K | ImageReward: 0.82 -> 0.95 | Improved |
| increase local interval number M from 4 to 8 (ablation) | LogCD (SDXL) · MSCOCO-2017 5K | ImageReward: 0.89 -> 0.95 | Improved |
| increase local interval number M beyond 8 (ablation) | LogCD (SDXL) · MSCOCO-2017 5K | ImageReward: 0.95 -> 0.98 | No clear effect |
| use multi-step teacher solver (p=3) in global CD instead of one step (ablation) | LogCD (SDXL) · MSCOCO-2017 5K | FID: 27.1 -> 25.9 | Improved |
| increase student denoising steps q from 2 to 3 (ablation) | LogCD (SDXL) · MSCOCO-2017 5K | ImageReward: 0.88 -> 0.94 | Improved |
| increase student denoising steps q beyond 3 (ablation) | LogCD (SDXL) · MSCOCO-2017 5K | CLIP score: 33.5 -> 33.6 | No clear effect |
| increase LogCM sampling steps from 2 to 4 | LogCM (SDXL) · MSCOCO-2017 5K | ImageReward: 0.93 -> 0.95 | Improved |
| distill FLUX.1-dev with LogCD instead of LCM (4 steps) | FLUX.1-dev · MSCOCO-2017 5K | CLIP score: 31.6 -> 32.6 | Improved |
| LogCM 4-step versus 25-step teacher on FLUX.1-dev | FLUX.1-dev · MSCOCO-2017 5K | CLIP score: 32.6 -> 32.6 | No clear effect |
Closest papers by meaning
- Invertible Consistency Distillation for Text-Guided Image Editing in Around 7 Steps (NeurIPS 2024 · similarity 0.89)
- SCott: Accelerating Diffusion Models with Stochastic Consistency Distillation (AAAI 2025 · similarity 0.88)
- Phased Consistency Models (NeurIPS 2024 · similarity 0.88)
- Hyper-SD: Trajectory Segmented Consistency Model for Efficient Image Synthesis (NeurIPS 2024 · similarity 0.88)
- Simplifying, Stabilizing and Scaling Continuous-time Consistency Models (ICLR 2025 · similarity 0.88)
- Improved Training Technique for Latent Consistency Models (ICLR 2025 · similarity 0.87)
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.
