Research map / Image generation / Text-to-image diffusion
CoDi: Subject-Consistent and Pose-Diverse Text-to-Image Generation
Zhanxin Gao, Beier Zhu, Liangyao, Jian Yang, Ying Tai
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 Identity Transport (IT) to vanilla SDXL (ablation) | SDXL · Subject-consistent T2I benchmark | DINO-v2 consistency: 0.8010 -> 0.8207 | Improved |
| add Identity Transport (IT) to vanilla SDXL (ablation) | SDXL · Subject-consistent T2I benchmark | pose diversity: 0.0772 -> 0.0800 | Improved |
| add Identity Refinement (IR) to vanilla SDXL (ablation) | SDXL · Subject-consistent T2I benchmark | DINO-v2 consistency: 0.8010 -> 0.8618 | Improved |
| add Identity Refinement (IR) alone to vanilla SDXL (ablation) | SDXL · Subject-consistent T2I benchmark | pose diversity: 0.0772 -> 0.0675 | Hurt |
| add IR after IT (full CoDi) versus IT only (ablation) | CoDi · Subject-consistent T2I benchmark | DINO-v2 consistency: 0.8207 -> 0.8514 | Improved |
| add IT before IR (full CoDi) versus IR only (ablation) | CoDi · Subject-consistent T2I benchmark | pose diversity: 0.0675 -> 0.0758 | Improved |
| add IT before IR (full CoDi) versus IR only (ablation) | CoDi · Subject-consistent T2I benchmark | DreamSim: 0.2044 -> 0.2136 | Hurt |
| increase alpha (top-percent reference features) from 30% to 70% (ablation) | CoDi · Subject-consistent T2I benchmark | DINO-v2 consistency / pose diversity | Mixed |
| vary stage transition step t from 2 to 30 (ablation) | CoDi · Subject-consistent T2I benchmark | DINO-v2 consistency / pose diversity | No clear effect |
| use CoDi instead of ConsiStory | ConsiStory · Subject-consistent T2I benchmark | DreamSim: 0.2336 -> 0.2136 | Improved |
| use CoDi instead of ConsiStory | ConsiStory · Subject-consistent T2I benchmark | pose diversity: 0.0621 -> 0.0758 | Improved |
| use CoDi instead of ConsiStory | ConsiStory · Subject-consistent T2I benchmark | prompt fidelity: 0.9148 -> 0.9041 | Hurt |
Closest papers by meaning
- Stable-Pose: Leveraging Transformers for Pose-Guided Text-to-Image Generation (NeurIPS 2024 · similarity 0.87)
- One-Prompt-One-Story: Free-Lunch Consistent Text-to-Image Generation Using a Single Prompt (ICLR 2025 · similarity 0.87)
- Storynizor: Consistent Story Generation via Inter-Frame Synchronized and Shuffled ID Injection (AAAI 2025 · similarity 0.86)
- XVerse: Consistent Multi-Subject Control of Identity and Semantic Attributes via DiT Modulation (NeurIPS 2025 · similarity 0.86)
- FaithfulFaces: Pose-Faithful Facial Identity Preservation for Text-to-Video Generation (NeurIPS 2026 · similarity 0.86)
- Storybooth: Training-Free Multi-Subject Consistency for Improved Visual Storytelling (ICLR 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.
