Research map / Image generation / Autoregressive and latent generation
ZigMa: A DiT-style Zigzag Mamba Diffusion Model
Vincent Tao Hu, Stefan A Baumann, Ming Gui, Olga Grebenkova, Pingchuan Ma, Johannes S Fischer, Bjorn Ommer
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 sweep scan with zigzag scan schemes (1 to 8) layer-wise (ablation) | Mamba diffusion backbone · MultiModal-CelebA-256 | FID 5k: 158.1 -> 45.5 | Improved |
| increase zigzag schemes from 1 to 8 (ablation) | ZigMa · MultiModal-CelebA-512 | FID 5k: 121.0 -> 34.9 | Improved |
| use Zigzag-8 instead of Zigzag-1 on text-conditioned MS-COCO (ablation) | ZigMa · MS-COCO 256 | FID 5k: 73.1 -> 41.8 | Improved |
| use ZigMa instead of VisionMamba | Mamba diffusion backbone · FacesHQ-1024 | FID 5k: 51.1 -> 37.8 | Improved |
| double the batch size (ablation) | ZigMa · FacesHQ-1024 | FID 5k: 37.8 -> 26.6 | Improved |
| add position embedding to ZigMa (ablation) | ZigMa · Unconditional CelebA 256 | FID: 14.27 -> 13.32 | No clear effect |
| use Hilbert space-filling curve instead of zigzag scan (ablation) | ZigMa · Unconditional MultiModal-CelebA256 | FID 5k: 13.32 -> 27.38 | Hurt |
| use in-context text concatenation instead of cross-attention (ablation) | ZigMa text conditioning · MultiModal-CelebA 256 | FID 5k: 45.5 -> 61.1 | Hurt |
| use full 3D zigzag scan instead of factorized 2D zigzag plus 1D sweep (ablation) | ZigMa video · UCF101 | FVD 5k: 210.2 -> 282.3 | Hurt |
| replace DiT or U-ViT backbone with ZigMa | unconditional diffusion backbone · Unconditional CelebA256 | memory (G): 29.20 -> 17.80 | Improved |
| increase patch size from 1 to 8 (ablation) | ZigMa · FID vs patch size | FID | Hurt |
| randomly shuffle patches instead of continuous scan (ablation) | ZigMa · MultiModal-CelebA | FID | Hurt |
Closest papers by meaning
- EfficientVMamba: Atrous Selective Scan for Light Weight Visual Mamba (AAAI 2025 · similarity 0.88)
- DiMSUM: Diffusion Mamba - A Scalable and Unified Spatial-Frequency Method for Image Generation (NeurIPS 2024 · similarity 0.88)
- Vision Mamba Mender (NeurIPS 2024 · similarity 0.88)
- Multi-Scale VMamba: Hierarchy in Hierarchy Visual State Space Model (NeurIPS 2024 · similarity 0.87)
- Scaling Diffusion Mamba with Bidirectional SSMs for Efficient 3D Shape Generation (AAAI 2025 · similarity 0.87)
- MobileMamba: Lightweight Multi-Receptive Visual Mamba Network (CVPR 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.
