Research map / Image generation / Fast diffusion sampling
SyncDiffusion: Coherent Montage via Synchronized Joint Diffusions
Yuseung Lee, Kunho Kim, Hyunjin Kim, Minhyuk Sung
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 |
|---|---|---|---|
| add perceptual-gradient synchronization using predicted denoised images | MultiDiffusion joint diffusion · Text-guided panorama generation (512×3072) | Intra-LPIPS: 0.69 ± 0.09 -> 0.56 ± 0.06 | Improved |
| add perceptual-gradient synchronization using predicted denoised images | MultiDiffusion joint diffusion · Text-guided panorama generation (512×3072) | Intra-Style-L: 2.98 ± 2.41 -> 1.39 ± 1.15 | Improved |
| add perceptual-gradient synchronization using predicted denoised images | MultiDiffusion joint diffusion · Text-guided panorama generation | human coherence preference: 33.65 -> 66.35 | Improved |
| increase gradient descent weight from 0 to 20 (ablation) | SyncDiffusion · Text-guided panorama generation (512×3072) | FID: 33.52 ± 12.43 -> 44.60 ± 18.45 | Hurt |
| increase gradient descent weight from 0 to 20 (ablation) | SyncDiffusion · Text-guided panorama generation (512×3072) | Mean-CLIP-S: 31.77 ± 2.32 -> 31.84 ± 2.19 | No clear effect |
| compute perceptual loss on noisy images rather than predicted denoised images (ablation) | SyncDiffusion · Text-guided panorama generation (512×3072) | Intra-LPIPS: 0.56 ± 0.06 -> 0.68 ± 0.09 | Hurt |
| use Style Loss instead of LPIPS as perceptual loss (ablation) | SyncDiffusion · Text-guided panorama generation (512×3072) | Intra-Style-L: 2.98 ± 2.41 -> 1.08 ± 1.10 | Improved |
| use Style Loss instead of LPIPS as perceptual loss (ablation) | SyncDiffusion · Text-guided panorama generation (512×3072) | FID: 33.52 ± 12.43 -> 73.05 ± 37.56 | Hurt |
| apply synchronization at 10 uniformly spaced steps rather than all 50 (ablation) | SyncDiffusion · Text-guided panorama generation (512×3072) | Intra-LPIPS: 0.69 ± 0.09 -> 0.62 ± 0.07 | Improved |
| apply synchronization only in the initial 3 denoising steps (ablation) | SyncDiffusion · Text-guided panorama generation (512×3072) | Intra-LPIPS: 0.69 ± 0.09 -> 0.62 ± 0.06 | Improved |
| apply synchronization only in the initial 3 denoising steps rather than at 10 uniform intervals (ablation) | SyncDiffusion · Text-guided panorama generation (512×3072) | FID: 36.22 ± 14.03 -> 35.40 ± 12.99 | Improved |
Closest papers by meaning
- StochSync: Stochastic Diffusion Synchronization for Image Generation in Arbitrary Spaces (ICLR 2025 · similarity 0.91)
- SyncTweedies: A General Generative Framework Based on Synchronized Diffusions (NeurIPS 2024 · similarity 0.91)
- ElasticDiffusion: Training-free Arbitrary Size Image Generation through Global-Local Content Separation (CVPR 2024 · similarity 0.88)
- Taming Stable Diffusion for Text to 360 Panorama Image Generation (CVPR 2024 · similarity 0.88)
- Move Anything with Layered Scene Diffusion (CVPR 2024 · similarity 0.88)
- Diffusing Differentiable Representations (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.
