Research map / 3D vision / 3D generation
Ar2Can: An Architect and an Artist Leveraging a Canvas for Multi-Human Generation
Shubhankar Borse, Phuc Pham, Farzad Farhadzadeh, Seokeon Choi, Phong Nguyen, Anh Tran, Sungrack Yun, Munawar Hayat, Fatih Porikli
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Hungarian centroid face-matching reward in place of simple exact-location matching (ablation) | Artist (Flux-Kontext GRPO) · MultiHuman-TestBench | Multi-ID similarity: 55.2 -> 60.3 | Improved |
| add Hungarian centroid face-matching reward in place of simple exact-location matching (ablation) | Artist (Flux-Kontext GRPO) · MultiHuman-TestBench | HPS: 27.6 -> 30.9 | Improved |
| add simple exact-location face-matching reward to Kontext (ablation) | Flux-Kontext · MultiHuman-TestBench | Count accuracy: 80.7 -> 75.6 | Hurt |
| add simple exact-location face-matching reward to Kontext (ablation) | Flux-Kontext · MultiHuman-TestBench | Multi-ID similarity: 14.5 -> 55.2 | Improved |
| add curriculum training over person count (ablation) | Artist with Hungarian reward · MultiHuman-TestBench | Count accuracy: 80.1 -> 86.9 | Improved |
| add curriculum training over person count (ablation) | Artist with Hungarian reward · MultiHuman-TestBench | Multi-ID similarity: 60.3 -> 68.2 | Improved |
| add curriculum training over person count (ablation) | Artist with Hungarian reward · MultiHuman-TestBench | HPS: 30.9 -> 30.8 | No clear effect |
| RL-train Artist with spatially-grounded rewards on Architect layouts | Flux-Kontext · MultiHuman-TestBench | Multi-ID similarity: 14.5 -> 68.2 | Improved |
| use Architect-A (Qwen) instead of Architect-B (Flux-Schnell) | Ar2Can · MultiHuman-TestBench | Count accuracy: 86.9 -> 90.2 | Improved |
| use Architect-A instead of Architect-B | Ar2Can · MultiHuman-TestBench | Action-C: 82.0 -> 77.6 | Hurt |
| Ar2Can vs WithAnyone on out-of-domain-biased MultiID-Test | WithAnyone · MultiID-Test | Multi-ID (Ref) similarity: 50.1 -> 54.3 | Improved |
| token dropping and sharing on canvas | Artist · Inference latency (A100, 3 identities) | latency | Improved |
Closest papers by meaning
- CapHuman: Capture Your Moments in Parallel Universes (CVPR 2024 · similarity 0.87)
- MultiHuman-Testbench: Benchmarking Image Generation for Multiple Humans (NeurIPS 2025 · similarity 0.87)
- ID-Patch: Robust ID Association for Group Photo Personalization (CVPR 2025 · similarity 0.86)
- FaithfulFaces: Pose-Faithful Facial Identity Preservation for Text-to-Video Generation (NeurIPS 2026 · similarity 0.86)
- Arc2Avatar: Generating Expressive 3D Avatars from a Single Image via ID Guidance (CVPR 2025 · similarity 0.85)
- When StyleGAN Meets Stable Diffusion: a W+ Adapter for Personalized Image Generation (CVPR 2024 · 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.
