Research map / Image generation / Autoregressive and latent generation
DDMI: Domain-agnostic Latent Diffusion Models for Synthesizing High-Quality Implicit Neural Representations
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DDMI, a latent diffusion model that generates adaptive positional embeddings (via hierarchical basis fields and coarse-to-fine conditioning) instead of INR network weights, enabling high-quality domain-agnostic INR generation across images, shapes, NeRF, and videos.
Problem
Existing domain-agnostic INR generative models generate MLP weights with fixed positional embeddings, limiting expressive power and yielding low-quality generation on large, complex datasets.
Approach
Generate sample-specific positional embeddings from latent-decoded basis fields rather than MLP weights; use an asymmetric VAE (D2C-VAE) to share a latent space between discrete data and continuous functions, add hierarchically decomposed multi-scale basis fields, and condition the MLP coarse-to-fine on these PEs; then train a latent diffusion model on that shared latent space.
Key results
- On CelebA-HQ 64^2, DDMI reaches FID 9.74 versus 13.2 for DPF, the best domain-agnostic baseline.
- On SkyTimelapse video, DDMI achieves FVD 66.25, beating StyleGAN-V (79.52) and DIGAN (83.11), and approaching PVDM (71.46).
- On ShapeNet multi-class, DDMI achieves MMD 1.3 and COV 0.421, the best MMD among the compared methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add HDBFs to the D2C-VAE baseline (ablation) | DDMI · AFHQv2 Cat 384² | PSNR: 32.72 -> 33.17 | Helped |
| Add HDBFs to the D2C-VAE baseline (ablation) | DDMI · AFHQv2 Cat 384² | FID: 8.54 -> 8.23 | Helped |
| Add CFC to the HDBF variant (ablation) | DDMI · AFHQv2 Cat 384² | PSNR: 33.17 -> 33.56 | Helped |
| Add CFC to the HDBF variant (ablation) | DDMI · AFHQv2 Cat 384² | FID: 8.23 -> 7.82 | Helped |
| Use 50 ODE function evaluations instead of 1000 (ablation) | DDMI · AFHQv2 Cat | FID: 4.27 -> 5.32 | Hurt |
| Use 25 ODE function evaluations instead of 1000 (ablation) | DDMI · AFHQv2 Cat | FID: 4.27 -> 5.93 | Hurt |
| Use 50 ODE function evaluations instead of 1000 (ablation) | DDMI · AFHQv2 Dog | FID: 8.54 -> 11.04 | Hurt |
| Use 25 ODE function evaluations instead of 1000 (ablation) | DDMI · AFHQv2 Dog | FID: 8.54 -> 14.23 | Hurt |
Limitations
- Video results are only competitive with, not better than, the discrete-pixel PVDM model on SkyTimelapse.
- Text-to-shape TMD (0.204) is worse than Diffusion-SDF (0.169).
Experimental setup
Datasets: AFHQv2 Cat AFHQv2 Dog CelebA-HQ CIFAR10 LSUN Churches ShapeNet Text2Shape SRN Cars SkyTimelapse
Benchmarks: CelebA-HQ AFHQv2 Cat AFHQv2 Dog CIFAR10 LSUN Churches ShapeNet Text2Shape SRN Cars SkyTimelapse
Metrics: FID precision recall MMD COV classification accuracy CLIP-S TMD FVD PSNR
Baselines: Functa GEM GASP DPF INR-GAN CIPS 3D-LDM SDF-Diffusion Diffusion-SDF DIGAN StyleGAN-V PVDM
Builds on: latent diffusion models Functa GASP CIPS PVDM
Closest papers by meaning
- Hyper-Transforming Latent Diffusion Models (ICML 2025 · similarity 0.89)
- Masked Autoencoders Are Effective Tokenizers for Diffusion Models (ICML 2025 · similarity 0.88)
- DisCo-Diff: Enhancing Continuous Diffusion Models with Discrete Latents (ICML 2024 · similarity 0.88)
- Reconstruction vs. Generation: Taming Optimization Dilemma in Latent Diffusion Models (CVPR 2025 · similarity 0.88)
- Neural Diffusion Models (ICML 2024 · similarity 0.87)
- Boosting Generative Image Modeling via Joint Image-Feature Synthesis (NeurIPS 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.