Research map / Image generation / Editing and style transfer
IMAGPose: A Unified Conditional Framework for Pose-Guided Person Generation
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 IMAGPose, a diffusion-based unified framework for pose-guided person generation that supports multiple target poses and multi-view source images via feature-level (VAE+image encoder), image-level (masked joint image/pose) conditioning, and cross-view attention.
Problem
Existing diffusion-based pose-guided person generation only maps one source image and one target pose to one target image, missing multi-target-pose and multi-source-image user scenarios and losing low-level texture detail due to frozen semantic encoders.
Approach
Combine low-level VAE latent texture tokens with high-level DINOv2 semantic tokens for appearance conditioning, inject source images directly as image-level conditions via a joint image/pose with a random masking strategy so one model handles any number of source or target images, and add a cross-view attention module over person images in the batch to keep local fidelity and global consistency.
Key results
- On DeepFashion 256x176, IMAGPose reaches SSIM 0.7561, LPIPS 0.1284, FID 5.8738, beating PCDMs (SSIM 0.7444, FID 7.4734) and PIDM (FID 6.3671).
- On Market-1501 128x64, IMAGPose reports SSIM 0.3282, LPIPS 0.2104, FID 12.659, outperforming PCDMs (FID 13.897).
- Inference is reported as nearly 8x faster than PIDM and nearly 3x faster than PoCoLD and CFLD, due to the masking strategy in ILC.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add CVA to B0 (ablation) | IMAGPose · DeepFashion | SSIM: 0.7235 -> 0.7281 | Helped |
| add FLC texture features to B0 (ablation) | IMAGPose · DeepFashion | SSIM: 0.7235 -> 0.7406 | Helped |
| add ILC image-level conditions to B1 (ablation) | IMAGPose · DeepFashion | SSIM: 0.7281 -> 0.7453 | Helped |
| add CVA to the model with FLC and ILC (ablation) | IMAGPose · DeepFashion | SSIM: 0.7528 -> 0.7561 | Helped |
| use multiple different source images (T3) instead of replicated source image (T2) | IMAGPose · DeepFashion | SSIM: 0.7594 -> 0.7727 | Helped |
| generate multiple target images with different poses at once (IMAGPose*) | IMAGPose · DeepFashion | SSIM | Mixed |
| replace the common single-image framework with M1 for multiple target poses | pose-guided generation · DeepFashion | SSIM: 0.7312 -> 0.7145 | Hurt |
| vary the image encoder and base diffusion model (ablation) | IMAGPose · DeepFashion | SSIM | No clear effect |
| replace Stable Diffusion V1.5 with Stable Diffusion V2.1 (ablation) | IMAGPose · DeepFashion | SSIM: 0.7561 -> 0.7572 | Mixed |
| add generated person images from IMAGPose to the re-identification training data | BoT · Market-1501 person re-identification, 20% real images | Rank1: 33.4 -> 66.4 | Helped |
Surprising findings
- A dedicated multi-target model (M1) performed worse than PIDM on SSIM (0.7145 vs 0.7312) and FID, despite supporting multiple target poses.
- Performance differences across base diffusion models and image encoders were small, suggesting the base model and encoder have minimal impact on results.
Limitations
- Supports generating at most four images at a time due to computational limits.
- Performs poorly on cartoon and non-photorealistic styles because training data is photorealistic.
- Speed comparisons and multi-view results are limited to DeepFashion and Market-1501 datasets.
Experimental setup
Datasets: DeepFashion Market-1501
Benchmarks: DeepFashion Market-1501
Metrics: SSIM LPIPS FID R2G G2R Jab (user preference) Rank-1 (person re-identification)
Baselines: PIDM PoCoLD CFLD PCDMs NTED CASD GFLA DPTN
Builds on: Stable Diffusion ControlNet PIDM DINOv2 BoT
Closest papers by meaning
- Advancing Pose-Guided Image Synthesis with Progressive Conditional Diffusion Models (ICLR 2024 · similarity 0.89)
- High-fidelity Person-centric Subject-to-Image Synthesis (CVPR 2024 · similarity 0.87)
- Coarse-to-Fine Latent Diffusion for Pose-Guided Person Image Synthesis (CVPR 2024 · similarity 0.87)
- Disentangled Pose and Appearance Guidance for Multi-Pose Generation (CVPR 2025 · similarity 0.86)
- MagicPose: Realistic Human Poses and Facial Expressions Retargeting with Identity-aware Diffusion (ICML 2024 · similarity 0.86)
- MultiHuman-Testbench: Benchmarking Image Generation for Multiple Humans (NeurIPS 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.