Research map / Image generation / Autoregressive and latent generation
Few-shot Implicit Function Generation via Equivariance
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.
Introduces Few-shot Implicit Function Generation and EQUIGEN, which generates diverse INR weights from few examples using a permutation-equivariant encoder, equivariance-guided diffusion, and subspace perturbation.
Problem
Generating diverse, functionally consistent INR weights from only a few target examples, where weight-space permutation symmetry breaks standard few-shot transfer and causes mode collapse.
Approach
Project INR weights into a permutation-equivariant latent space learned with SimCLR-style contrastive learning plus TSP-based weight smoothing, condition a transformer diffusion model on these features with an added equivariance loss, and generate diversity by Gaussian perturbations inside the equivariant subspace.
Key results
- On MNIST-INRs 10-shot, EQUIGEN reaches FID 121.24 and LPIPS 0.4133, versus 149.77 FID for the best prior baseline DAGAN.
- On ShapeNet-INRs 10-shot chair, EQUIGEN achieves MMD 4.2, COV 41%, 1-NNA 67.1%, improving over HyperDiffusion (MMD 7.2, COV 28%, 1-NNA 73.1%).
- On CIFAR-10-INRs 10-shot, EQUIGEN reaches FID 164.14 and LPIPS 0.4926, outperforming HyperDiffusion (186.58 FID, 0.3104 LPIPS).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use EQUIGEN instead of the strongest reported baseline | 10-shot INR generation · ShapeNet-INRs airplane | MMD: 5.0 -> 3.4 | Helped |
| use EQUIGEN instead of the strongest reported baseline | 10-shot INR generation · ShapeNet-INRs car | COV (%): 22 -> 31 | Helped |
| use EQUIGEN instead of the strongest reported baseline | 10-shot INR generation · ShapeNet-INRs chair | 1-NNA (%): 71.7 -> 67.1 | Helped |
| use EQUIGEN instead of the strongest reported baseline | 10-shot INR generation · MNIST-INRs | FID: 149.77 -> 121.24 | Helped |
| use EQUIGEN instead of the strongest reported baseline | 10-shot INR generation · CIFAR-10-INRs | LPIPS: 0.3104 -> 0.4926 | Helped |
| add equivariant conditioning rather than no condition (ablation) | EQUIGEN · MNIST-INRs | FID: 164.78 -> 121.24 | Helped |
| add equivariant conditioning rather than no condition (ablation) | EQUIGEN · MNIST-INRs | LPIPS: 0.2432 -> 0.4133 | Helped |
| replace class-label conditioning with equivariant conditioning (ablation) | EQUIGEN · MNIST-INRs | FID: 162.56 -> 121.24 | Helped |
| replace class-label conditioning with equivariant conditioning (ablation) | EQUIGEN · MNIST-INRs | LPIPS: 0.2437 -> 0.4133 | Helped |
| add weight-space smooth augmentation (ablation) | EQUIGEN · ShapeNet-INRs | generation quality and diversity | Helped |
| add equivariant subspace disturbance (ablation) | EQUIGEN · ShapeNet-INRs | generation diversity | Helped |
| add equivariant subspace disturbance (ablation) | EQUIGEN · ShapeNet-INRs | generation quality | Hurt |
Surprising findings
- Class-label conditioning gives only minimal improvement over the unconditional diffusion variant on MNIST-INRs (FID 162.56 vs 164.78), while equivariant conditioning gives a large gain (FID 121.24).
Limitations
- Diversity is traded against quality: larger subspace disturbance raises COV but also raises MMD.
- Gains from subspace disturbance are smaller without smooth augmentation, since the same disturbance can land in different category clusters.
- Evaluation relies on indirect metrics (rendered images, marching-cubes meshes) because no direct metric exists for generated weights.
Experimental setup
Datasets: MNIST-INRs CIFAR-10-INRs ShapeNet-INRs
Benchmarks: MNIST-INRs CIFAR-10-INRs ShapeNet-INRs
Metrics: FID LPIPS MMD COV 1-NNA
Baselines: FIGR DAGAN DAWSON INR2Vec HyperDiffusion PVD DPC Voxel Baseline
Builds on: SimCLR HyperDiffusion DDIM Equivariant architectures for deep weight spaces (NFN)
Closest papers by meaning
- VI^3NR: Variance Informed Initialization for Implicit Neural Representations (CVPR 2025 · similarity 0.89)
- Implicit Representations via Operator Learning (ICML 2024 · similarity 0.88)
- MIRE: Matched Implicit Neural Representations (CVPR 2025 · similarity 0.87)
- From Bricks to Bridges: Product of Invariances to Enhance Latent Space Communication (ICLR 2024 · similarity 0.86)
- Interpreting Equivariant Representations (ICML 2024 · similarity 0.86)
- Inductive Gradient Adjustment for Spectral Bias in Implicit Neural Representations (ICML 2025 · similarity 0.86)
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.