Research map / Graph learning / Graph self-supervised learning
Tracing Hyperparameter Dependencies for Model Parsing via Learnable Graph Pooling Network
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 LGPN, which casts model parsing (predicting 37 generative-model hyperparameters from an image) as graph node classification over a hyperparameter co-occurrence graph, with learnable pooling-unpooling and a high-resolution Generation Trace Capturing network.
Problem
Model parsing: predicting hyperparameters (loss functions, discrete and continuous architecture hyperparameters) of the generative model that produced a given image, while capturing dependencies among hyperparameters.
Approach
Represent hyperparameters as graph nodes (discrete nodes plus intervals for continuous values) with edges from label co-occurrence, and use a learnable pooling-unpooling GCN refinement block whose assignment matrices depend on the image features, combined with a high-resolution GTC backbone that avoids downsampling.
Key results
- LGPN reaches 84.6 F1 / 83.3 accuracy on loss function and 79.5 F1 / 77.5 accuracy on discrete architecture hyperparameters on RED140, outperforming FEN-PN (80.5 / 73.0 F1) and all baselines.
- On coordinated attack detection, LGPN achieves 85.9 accuracy, 95.7 AUC and 77.2 Pd@5% versus FEN-PN's 83.0, 92.4 and 61.2.
- Continuous architecture hyperparameter L1 error drops to 0.120 for LGPN versus 0.139 for FEN-PN on RED140.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace the MLP parsing head with GCN refinement (ablation) | GTC model parsing on RED140 · RED140 | Discrete architecture F1: 75.7 -> 79.5 | Helped |
| Replace the MLP parsing head with GCN refinement (ablation) | GTC model parsing on RED140 · RED140 | Continuous architecture L1 error: 0.135 -> 0.120 | Helped |
| Use GCN refinement rather than MLP layers (ablation) | Continuous architecture prediction · RED140, Param. Num. | L1 error: 0.148 -> 0.081 | Helped |
| Replace GCN refinement with stacked GCN (ablation) | LGPN model parsing · RED140, discrete architecture parameters | F1: 79.5 -> 76.9 | Hurt |
| Replace GCN refinement with Att-GCN (ablation) | LGPN model parsing · RED140, discrete architecture parameters | F1: 79.5 -> 78.0 | Hurt |
| Replace GCN refinement with Graph U-Net (ablation) | LGPN model parsing · RED140, discrete architecture parameters | F1: 79.5 -> 74.8 | Hurt |
| Add artifacts isolation loss (ablation) | LGPN training · RED140, discrete architecture parameters | F1: 1.1% | Helped |
| Add hyperparameter hierarchy loss (ablation) | LGPN training · RED140, discrete architecture parameters | F1: 2.0% | Helped |
| Use n=3 graph nodes per continuous hyperparameter (ablation) | LGPN continuous hyperparameter regression · Continuous hyperparameter regression | L1 error: 0.123 -> 0.120 | Helped |
| Increase correlation graph threshold to 0.65 (ablation) | LGPN · RED140 | Loss function F1: 84.6 -> 82.7 | Hurt |
Surprising findings
- FEN-PN, which performs competitively on RED116, only matches a simple HR-Net baseline on RED140, suggesting it degrades when diffusion models are included.
- Graph U-Net (which discards nodes when pooling) underperforms the simpler attention GCN on discrete architecture hyperparameters by 4.3%, indicating node dropping hurts model parsing.
Limitations
- Formulated as closed-set classification, so it cannot predict hyperparameters outside the predefined set (e.g., LeakyReLU variants).
- Evaluated only on the RED140 dataset and GMs within its scope; generalization to GMs outside that scope is untested.
Experimental setup
Datasets: RED116 RED140 CelebA MNIST CIFAR10 ImageNet ProGAN-generated images (CNN-generated detection) CelebA and LSUN (image attribution)
Benchmarks: RED140 RED116 CNN-generated image detection benchmark (11 unseen forgery methods) CelebA/LSUN image attribution
Metrics: F1 accuracy L1 error AUC Pd@5% FAR average precision
Baselines: FEN-PN ResNet-50 baseline HR-Net baseline ViT-B baseline Att-GCN Graph U-Net HiFi-Net NPR
Builds on: Graph U-Net graph attention networks FEN-PN HiFi-Net
Closest papers by meaning
- Graph Parsing Networks (ICLR 2024 · similarity 0.87)
- GRAVER: Generative Graph Vocabularies for Robust Graph Foundation Models Fine-tuning (NeurIPS 2025 · similarity 0.85)
- Deep Biaffine Attention for Neural Dependency Parsing (ICLR 2017 · similarity 0.84)
- Generative Graph Pattern Machine (NeurIPS 2025 · similarity 0.84)
- On the Role of Edge Dependency in Graph Generative Models (ICML 2024 · similarity 0.84)
- Mirage: Model-agnostic Graph Distillation for Graph Classification (ICLR 2024 · similarity 0.83)
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.