Research map / Image generation / Autoregressive and latent generation
Moebius: 0.2B Lightweight Image Inpainting Framework with 10B-Level Performance
Kangsheng Duan, Ziyang Xu, Wenyu Liu, Xiaohu Ruan, Xiaoxin Chen, Xinggang Wang
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 knowledge distillation (ablation) | Moebius architecture · Places2 (Test) | FID: 33.42 -> 26.43 | Improved |
| add knowledge distillation (ablation) | L-lambda + I-lambda Conv model · Places2 (Test) | FID: 33.21 -> 24.73 | Improved |
| add fine-grained distillation, task loss and latent perceptual loss to coarse-only distillation (ablation) | Moebius · Places2 (Test) | FID: 74.20 -> 26.43 | Improved |
| replace standard convolutions with depthwise convolutions, no distillation (ablation) | teacher-style architecture · Places2 (Test) | FID: 32.75 -> 43.58 | Hurt |
| replace GLA with Local-lambda only, no distillation (ablation) | teacher-style architecture · Places2 (Test) | FID: 32.75 -> 37.65 | Hurt |
| replace cross-attention with Interactive-lambda only, no distillation (ablation) | teacher-style architecture · Places2 (Test) | FID: 32.75 -> 36.91 | Hurt |
| replace FFN with MixFFN only, no distillation (ablation) | teacher-style architecture · Places2 (Test) | FID: 32.75 -> 35.24 | Hurt |
| combine Local-lambda and Interactive-lambda, no distillation (ablation) | teacher-style architecture · Places2 (Test) | FID: 32.75 -> 33.21 | No clear effect |
| use DWConv instead of Conv with distillation (ablation) | L-lambda-I-lambda-FFN · Places2 (Test) | FID: 24.73 -> 25.86 | Mixed |
| add MixFFN to DWConv model with distillation (ablation) | L-lambda-I-lambda · Places2 (Test) | FID: 25.86 -> 26.43 | Mixed |
| isolated lightweight operators under distillation versus teacher-style with distillation (ablation) | teacher-style architecture · Places2 (Test) | FID: 26.81 -> 29.41 | Hurt |
| tune CFG scale to 2.5 (ablation) | Moebius · Places2 (Test) | FID: 10.14 -> 9.48 | Improved |
Closest papers by meaning
- Image Inpainting via Tractable Steering of Diffusion Models (ICLR 2024 · similarity 0.85)
- HD-Painter: High-Resolution and Prompt-Faithful Text-Guided Image Inpainting with Diffusion Models (ICLR 2025 · similarity 0.85)
- Don't Look into the Dark: Latent Codes for Pluralistic Image Inpainting (CVPR 2024 · similarity 0.85)
- PixPerfect: Seamless Latent Diffusion Local Editing with Discriminative Pixel-Space Refinement (NeurIPS 2025 · similarity 0.85)
- Reconstructing the Image Stitching Pipeline: Integrating Fusion and Rectangling into a Unified Inpainting Model (NeurIPS 2024 · similarity 0.85)
- Towards Enhanced Image Inpainting: Mitigating Unwanted Object Insertion and Preserving Color Consistency (CVPR 2025 · 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.
