Research map / Image generation / Autoregressive and latent generation
GRPose: Learning Graph Relations for Human Image Generation with Pose Priors
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 GRPose, which builds graph structures over pose priors and diffusion latents with a Progressive Graph Integrator in a ControlNet-style adapter, plus a pose perception loss, for pose-guided human image generation on Human-Art and LAION-Human.
Problem
Pose-guided human image generation with diffusion models often yields poor alignment with pose priors, since existing methods inject pose information in Euclidean space and fail to model higher-order relations between pose parts.
Approach
Treat spatial points of the encoded pose prior and the latent representation as graph nodes connected by kNN edges, apply graph convolutions hierarchically inside each Adapter layer (Progressive Graph Integrator), and add a pose perception loss computed from a pretrained pose estimator's features.
Key results
- On Human-Art, GRPose reaches 49.50% AP and 70.84% SAP, versus 39.52% AP and 69.19% SAP for ControlNet and 44.57% AP for HumanSD.
- On LAION-Human, GRPose reaches 57.01% AP versus 50.95% for HumanSD, with PCE 1.29 versus 1.25 for HumanSD (best PCE there).
- Ablation on Human-Art: PGI alone raises ControlNet AP from 39.52 to 47.80; pose perception loss alone to 43.81; both together 49.50.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use GRPose instead of HumanSD | Human-Art generation · Human-Art | AP (%): 44.57 -> 49.50 | Helped |
| use GRPose instead of HumanSD | LAION-Human generation · LAION-Human | AP (%): 50.95 -> 57.01 | Helped |
| add PGI (ablation) | ControlNet on Human-Art · Human-Art | AP (%): 39.52 -> 47.80 | Helped |
| add PGI (ablation) | ControlNet on Human-Art · Human-Art | SAP (%): 69.19 -> 70.63 | Helped |
| add pose perception loss (ablation) | ControlNet on Human-Art · Human-Art | AP (%): 39.52 -> 43.81 | Helped |
| add pose perception loss (ablation) | ControlNet on Human-Art · Human-Art | SAP (%): 69.19 -> 70.37 | Helped |
| increase graph layer count from 0 to 3 (ablation) | GRPose on Human-Art · Human-Art | AP (%): 41.85 -> 47.80 | Helped |
| set pose loss weight α to 0.1 instead of 0.01 (ablation) | GRPose on Human-Art · Human-Art | AP (%): 49.50 -> 45.58 | Hurt |
| set pose loss weight α to 0.05 instead of 0.01 (ablation) | GRPose on Human-Art · Human-Art | AP (%): 49.50 -> 48.44 | Hurt |
| use GRPose instead of ControlNet | Human-Art cartoon subset, upper-body generation · Human-Art cartoon subset, upper body | AP (%): 15.51 -> 31.87 | Helped |
| use GRPose instead of ControlNet | Human-Art cartoon subset, full-body generation · Human-Art cartoon subset, full body | AP (%): 30.78 -> 39.19 | Helped |
Surprising findings
- On Human-Art, GRPose has higher FID (13.76) and KID than ControlNet and HumanSD, and on LAION-Human its PCE is slightly worse than HumanSD's, so pose gains come with some image-quality trade-offs.
Limitations
- Graph-stage ablation was limited to 4 layers due to computational resource constraints.
- Image quality metrics (FID/KID, CLIP-Score) slightly decline relative to some baselines.
- The paper gives little detail on the pose estimator used in training versus evaluation and no human evaluation.
Experimental setup
Datasets: Human-Art LAION-Human
Benchmarks: Human-Art LAION-Human
Metrics: AP SAP PCE FID KID CLIP-Score
Baselines: SD1.5 T2I-Adapter ControlNet Uni-ControlNet HumanSD
Builds on: ControlNet HumanSD Stable Diffusion Graph Convolutional Networks DDIM
Closest papers by meaning
- Advancing Pose-Guided Image Synthesis with Progressive Conditional Diffusion Models (ICLR 2024 · similarity 0.86)
- FlexPose: Pose Distribution Adaptation with Limited Guidance (AAAI 2025 · similarity 0.86)
- Disentangled Pose and Appearance Guidance for Multi-Pose Generation (CVPR 2025 · similarity 0.86)
- Stable-Pose: Leveraging Transformers for Pose-Guided Text-to-Image Generation (NeurIPS 2024 · similarity 0.86)
- $\text{Di}^2\text{Pose}$: Discrete Diffusion Model for Occluded 3D Human Pose Estimation (NeurIPS 2024 · similarity 0.85)
- Can Generative Video Models Help Pose Estimation? (CVPR 2025 · similarity 0.85)
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.