Research map / Image generation / Autoregressive and latent generation
Controllable Data Generation with Hierarchical 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 CHINR, a two-stage generative implicit neural representation framework that uses a Layers-of-Experts INR with per-layer latents and a hierarchical conditional diffusion model to enable layer-wise controllable generation.
Problem
Generative implicit neural representations model INR parameter distributions with flat latents, ignoring the hierarchical structure of INR layers and data semantics, which limits controllability and generalization.
Approach
Configure the INR as a Layers-of-Experts network where each layer has shared expert weights modulated by a per-instance latent, then model layer-wise latents with a diffusion model conditioned on binarized previous-layer latents, giving a coarse-to-fine controllable generation chain.
Key results
- CHINR reaches 34.9 PSNR and 0.964 SSIM on CelebA-HQ reconstruction, versus 34.5 PSNR for mNIF.
- On ShapeNet, CHINR achieves 22.3 PSNR, 0.988 accuracy and 0.441 coverage, the best among compared methods.
- On CelebA-HQ generation CHINR has FID 13.4, slightly behind mNIF's 13.2, but a higher distance-to-training-set score (15.971 vs 6.243) indicating less memorization.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CHINR instead of mNIF | CelebA-HQ reconstruction · CelebA-HQ | PSNR: 34.5±0.2 -> 34.9±0.3 | Helped |
| Use CHINR instead of mNIF | CelebA-HQ reconstruction · CelebA-HQ | SSIM: 0.957±0.005 -> 0.964±0.006 | Helped |
| Use CHINR instead of mNIF | CelebA-HQ generation · CelebA-HQ | FID: 13.2±0.1 -> 13.4±0.1 | Hurt |
| Use CHINR instead of Functa | ShapeNet reconstruction · ShapeNet | PSNR: 22.1±0.3 -> 22.3±0.2 | Helped |
| Use CHINR instead of Functa | ShapeNet reconstruction · ShapeNet | Accuracy: 0.983±0.005 -> 0.988±0.005 | Helped |
| Use CHINR instead of mNIF | ShapeNet generation · ShapeNet | Coverage: 0.435±0.003 -> 0.441±0.002 | Helped |
| Use CHINR instead of mNIF | SRN-Cars reconstruction · SRN-Cars | PSNR: 26.0±0.3 -> 26.3±0.2 | Helped |
| Use CHINR instead of mNIF | SRN-Cars reconstruction · SRN-Cars | SSIM: 0.763±0.013 -> 0.780±0.011 | Helped |
| Use CHINR instead of mNIF | SRN-Cars generation · SRN-Cars | FID: 79.3±0.3 -> 77.8±0.2 | Helped |
| Add conditional modeling at layer 5 (ablation) | CHINR generation on CelebA-HQ · CelebA-HQ | FID: 13.6 -> 13.4 | Helped |
| Add conditional modeling at layers 4 and 5 (ablation) | CHINR generation on CelebA-HQ · CelebA-HQ | FID: 15.5 -> 13.4 | Helped |
| Add conditional modeling at layers 3, 4, and 5 (ablation) | CHINR generation on CelebA-HQ · CelebA-HQ | FID: 52.8 -> 13.4 | Helped |
Surprising findings
- On CelebA-HQ, CHINR's FID is slightly worse than mNIF's, while the paper argues mNIF memorizes the training set more.
Limitations
- Scalability to larger and more diverse datasets is limited because the shared network may struggle to capture complex patterns.
- Latent composition can break the conditional chain and cause artifacts.
- Binary condition length must be tuned; too short loses conditional dependency, too long makes the chain near one-to-one and reduces controllability.
Experimental setup
Datasets: CelebA-HQ 64 ShapeNet 64^3 SRN-Cars AMASS
Benchmarks: CelebA-HQ 64 ShapeNet 64^3 SRN-Cars
Metrics: PSNR SSIM accuracy FID coverage
Baselines: Functa GEM GASP mNIF
Builds on: Functa mNIF CAVIA SIREN Denoising diffusion probabilistic models
Closest papers by meaning
- Hyper-Transforming Latent Diffusion Models (ICML 2025 · similarity 0.86)
- Banyan: Improved Representation Learning with Explicit Structure (ICML 2025 · similarity 0.86)
- Layer-Wise Representation Fusion for Compositional Generalization (AAAI 2024 · similarity 0.85)
- Few-shot Implicit Function Generation via Equivariance (CVPR 2025 · similarity 0.84)
- Neural-Symbolic Recursive Machine for Systematic Generalization (ICLR 2024 · similarity 0.84)
- Nested Diffusion Models Using Hierarchical Latent Priors (CVPR 2025 · similarity 0.84)
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.