Research map / Image generation / Text-to-image diffusion
MagicQuill V2: Precise and Interactive Image Editing with Layered Visual Cues
Zichen Liu, Yue Yu, Hao Ouyang, Qiuyu Wang, Shuailei Ma, Ka Leong Cheng, Wen Wang, Qingyan Bai, Yuxuan Zhang, Yanhong Zeng, Yixuan Li, Xing Zhu and 2 more
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 edge layer and color layer together instead of edge only (ablation) | MagicQuill V2 control layers · Pico-Banana-400K, 1,000 cases | LPIPS: 0.317 -> 0.283 | Improved |
| add color layer to the edge layer (ablation) | MagicQuill V2 control layers · Pico-Banana-400K, 1,000 cases | L1: 0.107 -> 0.080 | Improved |
| add edge layer to the color layer (ablation) | MagicQuill V2 control layers · Pico-Banana-400K, 1,000 cases | L1: 0.080 -> 0.080 | No clear effect |
| use MagicQuill V2 edge layer instead of Qwen-Image edge control module | edge-guided editing · Pico-Banana-400K, 1,000 cases | L1: 0.131 -> 0.107 | Improved |
| add the data-construction pipeline (completion LoRA, relight, perspective, resolution augmentations) to FLUX Kontext | content layer · Content layer test set (200 samples) | LPIPS: 0.359 -> 0.202 | Improved |
| remove perspective augmentation (ablation) | content layer training data · qualitative cabinet example | qualitative geometry correction | Hurt |
| remove relight augmentation (ablation) | content layer training data · qualitative bar example | qualitative lighting harmony | Hurt |
| remove object completion LoRA from data pipeline (ablation) | content layer training data · qualitative occluded book example | qualitative interaction | Hurt |
| remove resolution augmentation (ablation) | content layer training data · qualitative low-quality cues | qualitative image quality | Hurt |
| increase control strength sigma (ablation) | edge and color layers · qualitative control strength analysis | cue adherence | Mixed |
| fine-tune spatial branch for removal and use MagicQuill V2 instead of OmniEraser (ControlNet) | object removal · RORD, 5,000 samples | FID: 25.92 -> 16.42 | Improved |
Closest papers by meaning
- MagicQuill: An Intelligent Interactive Image Editing System (CVPR 2025 · similarity 0.91)
- Masked Region Transformer for Layered Image Generation and Editing at Scale (CVPR 2026 · similarity 0.88)
- Generative Image Layer Decomposition with Visual Effects (CVPR 2025 · similarity 0.88)
- LayerCraft: Enhancing Text-to-Image Generation with CoT Reasoning and Layered Object Integration (NeurIPS 2025 · similarity 0.88)
- Imagine Before You Draw: Visual Prompt Engineering for Image Generation (NeurIPS 2026 · similarity 0.88)
- Textualize Visual Prompt for Image Editing via Diffusion Bridge (AAAI 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.
