Research map / Image generation / Autoregressive and latent generation
Reviving ConvNeXt for Efficient Convolutional Diffusion Models
Taesung Kwon, Lorenzo Bianchi, Lennart Wittke, Felix Watine, Fabio Carrara, Jong Chul Ye, Romann Weber, Vinicius Azevedo
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 |
|---|---|---|---|
| use FCDM-XL instead of DiT-XL/2 | latent diffusion, ImageNet 256x256, no guidance · ImageNet 256x256 | FID: DiT-XL/2 400K 19.47 -> FCDM-XL 400K 10.72 | Improved |
| use FCDM-XL at 1M iterations vs DiT-XL/2 at 7M | latent diffusion, ImageNet 256x256 · ImageNet 256x256 | FID: 9.62 -> 7.91 | Improved |
| use FCDM backbone instead of DiT at equal parameter count | all scales S/B/L/XL · ImageNet 256x256 | FLOPs (G): XL 118.6 -> 64.6 | Improved |
| use FCDM-L instead of DiCo-L | L scale, 400K iters · ImageNet 256x256 | FID: 13.66 -> 13.83 | No clear effect |
| use FCDM-XL at 512x512 | latent diffusion, ImageNet 512x512 · ImageNet 512x512 | FID: DiT-XL/2 3M 12.03 -> FCDM-XL 1M 7.46 | Improved |
| shrink depthwise kernel from 7x7 (ablation) | FCDM-L · ImageNet 256x256 200K iters | FID: 19.97 -> 20.48 (5x5), 21.28 (3x3) | Hurt |
| replace GRN with compact channel attention from DiCo (ablation) | FCDM-L · ImageNet 256x256 200K iters | FID: 19.97 -> 23.85 | Hurt |
| add DiCo-style feedforward module (ablation) | FCDM-L · ImageNet 256x256 200K iters | FID: 19.97 -> 28.52 | Hurt |
| add inverted bottleneck (ablation) | FCDM-L · ImageNet 256x256 200K iters | FID: 28.76 -> 19.97 | Improved |
| use FCDM block instead of ResNet block (ablation) | FCDM-L · ImageNet 256x256 200K iters | FID: 31.14 -> 19.97 | Improved |
| scale model size from S to XL (ablation) | FCDM · ImageNet 256x256 400K iters | FID: 48.52 -> 26.21 -> 13.83 -> 10.72 (S, B, L, XL) | Improved |
| use FCDM-XL with guidance instead of latest state of the art (EDM-2, Simpler Diffusion) | FCDM-XL · ImageNet 256x256 | FID | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- DiCo-XL (400K) — DiCo: Revitalizing ConvNets for Scalable and Efficient Diffusion Modeling (NeurIPS 2025) · 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
- DiCo: Revitalizing ConvNets for Scalable and Efficient Diffusion Modeling (NeurIPS 2025 · similarity 0.92)
- DiC: Rethinking Conv3x3 Designs in Diffusion Models (CVPR 2025 · similarity 0.89)
- Neural Diffusion Models (ICML 2024 · similarity 0.89)
- LaVin-DiT: Large Vision Diffusion Transformer (CVPR 2025 · similarity 0.88)
- Learning Stackable and Skippable LEGO Bricks for Efficient, Reconfigurable, and Variable-Resolution Diffusion Modeling (ICLR 2024 · similarity 0.88)
- DiG: Scalable and Efficient Diffusion Models with Gated Linear Attention (CVPR 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.
