Research map / Image generation / Text-to-image diffusion
Referring Object Manipulation of Natural Images with Conditional Classifier-Free Guidance
Myungsub Choi
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 |
|---|---|---|---|
| replace classifier-free guidance with conditional classifier-free guidance starting from the referring prompt (ablation) | GLIDE inpainting with MDETR masks · User study, average rank | average rank (lower better): 2.57 -> 2.20 | Improved |
| replace classifier-free guidance with conditional classifier-free guidance (ablation) | GLIDE inpainting with MDETR masks · User study vs GLIDE (CF-guided) | winning probability: 58.4% | Improved |
| use conditional CFG instead of CLIP guidance (ablation) | GLIDE inpainting with MDETR masks · User study vs GLIDE (CLIP-guided) | winning probability: 2.82 -> 2.20 | Improved |
| use conditional CFG instead of Blended diffusion with same masks | Referring object manipulation framework · User study vs Blended | winning probability: 2.40 -> 2.20 | Improved |
| add localized ranking on 1.3x enlarged bounding box instead of masked-region ranking (ablation) | Output ranking of generated samples · qualitative, PhraseCut | visual realism (qualitative) | Improved |
| add mask dilation to the predicted segmentation mask (k=5 vs k=1 or 3) (ablation) | Inpainting with predicted masks · qualitative, PhraseCut | artifact avoidance (qualitative) | Improved |
| use larger guidance scale s = 15 instead of s = 5 for conditional CFG (ablation) | Conditional CFG · qualitative | edit quality: 5 -> 15 | Improved |
| condition on the original unmasked image with referring prompt instead of masked-out input (ablation) | Conditional CFG vs CF-guided GLIDE · qualitative | preservation of original characteristics | Mixed |
Closest papers by meaning
- Referring Image Editing: Object-level Image Editing via Referring Expressions (CVPR 2024 · similarity 0.90)
- Multi-Region Text-Driven Manipulation of Diffusion Imagery (AAAI 2024 · similarity 0.88)
- An Item Is Worth a Prompt: Versatile Image Editing with Disentangled Control (AAAI 2025 · similarity 0.88)
- Prompt Augmentation for Self-supervised Text-guided Image Manipulation (CVPR 2024 · similarity 0.87)
- Add-it: Training-Free Object Insertion in Images With Pretrained Diffusion Models (ICLR 2025 · similarity 0.86)
- Early Semantic Grounding in Image Editing Models for Zero-Shot Referring Image Segmentation (NeurIPS 2026 · 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.
