Research map / Image generation / Autoregressive and latent generation
Group Diffusion: Enhancing Image Generation by Unlocking Cross-Sample Collaboration
Sicheng Mo, Thao Nguyen, Richard Zhang, Nick Kolkin, Siddharth Srinivasan Iyer, Eli Shechtman, Krishna Kumar Singh, Yong Jae Lee, Bolei Zhou, Yuheng Li
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 group attention on unconditional model only (C=1, UC=4, class query) (ablation) | DiT-XL/2 · ImageNet 256x256 | FID (w/ CFG): 3.08 -> 2.81 | Improved |
| increase group size from 2 to 16 (CLIP-L) (ablation) | GroupDiff DiT-XL/2 · ImageNet 256x256 | FID (w/ CFG): 2.92 -> 1.86 | Improved |
| replace random groups with CLIP-L similarity groups (ablation) | GroupDiff DiT-XL/2 · ImageNet 256x256 | FID (w/ CFG): 3.57 -> 2.42 | Improved |
| use class-based instead of random groups (ablation) | GroupDiff DiT-XL/2 · ImageNet 256x256 | FID (w/ CFG): 3.57 -> 2.81 | Improved |
| swap pretrained encoder for query (CLIP-L vs DINOv2-B) (ablation) | GroupDiff DiT-XL/2 · ImageNet 256x256 | FID (w/ CFG): 2.42 vs 2.51 | No clear effect |
| add noise-level variation of 50 within group (ablation) | GroupDiff DiT-XL/2 · ImageNet 256x256 | FID (w/ CFG): 2.42 -> 2.34 | Improved |
| apply group attention only in timesteps 0.0-0.4 (ablation) | GroupDiff DiT-XL/2 · ImageNet 256x256 | FID-10K: 4.21 -> 3.92 | Improved |
| apply group attention up to timestep 0.6 (ablation) | GroupDiff DiT-XL/2 · ImageNet 256x256 | FID-10K: 4.21 -> 4.63 | Hurt |
| keep group attention in layers 1-9 (ablation) | GroupDiff DiT-XL/2 · ImageNet 256x256 | FID-10K: 294.38 -> 4.21 | Improved |
| keep group attention in layers 20-27 (ablation) | GroupDiff DiT-XL/2 · ImageNet 256x256 | FID-10K: 4.49 -> 4.21 | No clear effect |
| use CLIP-L instead of class query when resuming SiT-XL/2 for 100 epochs (ablation) | SiT-XL/2 GroupDiff-4* · ImageNet 256x256 | FID: 1.76 -> 1.40 | Improved |
| add GroupDiff-4 trained from scratch to DiT-XL/2 | DiT-XL/2 · ImageNet 256x256 | FID: 2.27 -> 1.66 | Improved |
Closest papers by meaning
- Faster Diffusion: Rethinking the Role of the Encoder for Diffusion Model Inference (NeurIPS 2024 · similarity 0.89)
- Tracing the Roots: Leveraging Temporal Dynamics in Diffusion Trajectories for Origin Attribution (NeurIPS 2025 · similarity 0.89)
- A Simple Approach to Unifying Diffusion-based Conditional Generation (ICLR 2025 · similarity 0.89)
- CCS: Controllable and Constrained Sampling with Diffusion Models via Initial Noise Perturbation (NeurIPS 2025 · similarity 0.89)
- Learning Diffusion Models with Flexible Representation Guidance (NeurIPS 2025 · similarity 0.89)
- DiffuseHigh: Training-Free Progressive High-Resolution Image Synthesis Through Structure Guidance (AAAI 2025 · 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.
