Research map / Image generation / Autoregressive and latent generation
Diffusion Image Generation with Explicitly Modeling of Data Manifold Geometry
Duoduo Xue, Zhiyu Zhu, Junhui Hou
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 DiT-B/2 with MIND-B (discrete-token hypersphere diffusion) | DiT-B/2 · ImageNet 256x256 (80 epochs, no guidance) | FID: 43.47 -> 22.73 | Improved |
| replace DiT-S/2 with MIND-S | DiT-S/2 · ImageNet 256x256 (80 epochs, no guidance) | FID: 68.40 -> 40.72 | Improved |
| replace SiT-B/2 with MIND-B | SiT-B/2 · ImageNet 256x256 (80 epochs, no guidance) | FID: 33.02 -> 22.73 | Improved |
| replace SiT-B/2 with MIND-B | SiT-B/2 · ImageNet 256x256 (80 epochs, no guidance) | Recall: 0.63 -> 0.58 | Hurt |
| replace DiT-S/2 with MIND-S under guidance cfg 1.5 | DiT-S/2 · ImageNet 256x256 (80 epochs, cfg 1.5) | FID: 44.42 -> 22.33 | Improved |
| replace SiT-S/2 with MIND-S under guidance cfg 1.5 | SiT-S/2 · ImageNet 256x256 (80 epochs, cfg 1.5) | FID: 35.62 -> 22.33 | Improved |
| replace DiT-B/2 with MIND-B under guidance cfg 2.0 | DiT-B/2 · ImageNet 256x256 (80 epochs, cfg 2.0) | FID: 10.95 -> 7.97 | Improved |
| replace SiT-B/2 with MIND-B under guidance cfg 2.0 | SiT-B/2 · ImageNet 256x256 (80 epochs, cfg 2.0) | FID: 9.07 -> 7.97 | Improved |
| replace DiT-XL/2 with MIND-XL-G in system-level comparison | DiT-XL/2 · ImageNet 256x256 (system-level, with guidance) | FID: 2.27 -> 1.95 | Improved |
| use stronger GigaTok tokenizer and longer training for MIND-B | MIND-B · ImageNet 256x256 (system-level, with guidance) | FID: 2.18 -> 2.06 | Improved |
| use MIND-B (130M) instead of LlamaGen-3B (3.1B) | LlamaGen-3B · ImageNet 256x256 (system-level, with guidance) | FID: 2.18 -> 2.06 | Improved |
| use MIND-S instead of eMIGM-S at one-third of the parameters | eMIGM-S · ImageNet 256x256 (80 epochs, no guidance) | FID: 44.47 -> 40.72 | Improved |
Closest papers by meaning
- Manifold Preserving Guided Diffusion (ICLR 2024 · similarity 0.91)
- What Matters for Diffusion-Friendly Latent Manifold? Prior-Aligned Autoencoders for Latent Diffusion (NeurIPS 2026 · similarity 0.90)
- Generative Modeling on Manifolds Through Mixture of Riemannian Diffusion Processes (ICML 2024 · similarity 0.90)
- Diffusion Models Encode the Intrinsic Dimension of Data Manifolds (ICML 2024 · similarity 0.90)
- Manifolds, Random Matrices and Spectral Gaps: The geometric phases of generative diffusion (ICLR 2025 · similarity 0.89)
- What Secrets Do Your Manifolds Hold? Understanding the Local Geometry of Generative Models (ICLR 2025 · similarity 0.89)
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.
