Research map / 3D vision / 3D generation
GeoRemover: Removing Objects and Their Causal Visual Artifacts
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Decouples object removal into depth-map geometry removal (diffusion with DPO-style flow preference loss) and geometry-conditioned RGB rendering, removing objects along with their shadows and reflections.
Problem
Object removal should also eliminate causal visual artifacts (shadows, reflections) that lie outside the user mask, while avoiding over-erasing other objects; existing mask-aligned and loosely mask-aligned methods either miss these artifacts or lack controllability.
Approach
Remove the object in the depth (geometry) domain under strictly mask-aligned supervision, using a Bradley-Terry preference loss over smooth vs. sharp depth-flow completions to suppress hallucinated structures; then render the RGB image conditioned on the before/after depth maps so that causal effects vanish implicitly.
Key results
- On RORD-Val, the method reaches FID 31.15 and PSNR 23.70 versus OmniEraser's 43.71 and 22.13.
- On RemovalBench, FID 29.88 and CMMD 0.089 vs. OmniEraser's 39.52 and 0.208.
- On the authors' CausRem artifact benchmark, IoU of residue removal is 73.76% vs. OmniEraser's 68.29%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace one-stage editing with two-stage geometry removal and appearance rendering (ablation) | GeoRemover · RORD-Val | FID: 56.24 -> 31.15 | Helped |
| Add DPO supervision to the two-stage method (ablation) | GeoRemover · RORD-Val | Insert.: 5.09% -> 1.48% | Helped |
| Add DPO supervision to the two-stage method (ablation) | GeoRemover · RORD-Val | MAE: 0.0490 -> 0.0387 | Helped |
| Use bidirectional rather than unidirectional rendering (ablation) | GeoRemover · RORD-Val | FID: 38.43 -> 31.15 | Helped |
| Use bidirectional rather than unidirectional rendering (ablation) | GeoRemover · RORD-Val | AS: 4.19 -> 4.69 | Helped |
| Use geometry-aware two-stage removal and rendering | GeoRemover · RemovalBench | FID: 39.52 -> 29.88 | Helped |
| Use geometry-aware two-stage removal and rendering | GeoRemover · RORD-Val | LPIPS: 0.166 -> 0.103 | Helped |
| Use geometry-aware two-stage removal and rendering | GeoRemover · CausRem | IoU%: 68.29 -> 73.76 | Helped |
Surprising findings
- The one-stage depth-conditioned variant underperforms the two-stage model, indicating the decoupling itself (not just depth input) drives gains.
- Stage 2 may hallucinate diffuse glow for self-emitting objects such as colored bulbs rather than removing their illumination.
Limitations
- Stage 1 fails when depth within the mask is unchanged (motion blur, transparency, specular surfaces), requiring a heuristic Local Max Depth Fill-in.
- Incomplete masks cause Stage 1 to hallucinate extra objects.
- Self-emitting and transparent objects can yield residual glow or color bleeding.
Experimental setup
Datasets: RORD RORD-Val RemovalBench
Benchmarks: RemovalBench RORD-Val CausRem
Metrics: FID CMMD LPIPS PSNR AS IoU MAE SSIM DISTS DreamSim
Baselines: ZITS++ MAT LaMa RePaint BLD LDM SD-Inpaint SDXL-Inpaint BrushNet FLUX.1-Fill PowerPaint CLIPAway
Builds on: Direct Preference Optimization FLUX.1-Fill Depth Anything LoRA RORD
Closest papers by meaning
- SmartEraser: Remove Anything from Images using Masked-Region Guidance (CVPR 2025 · similarity 0.86)
- In-N-Out: Lifting 2D Diffusion Prior for 3D Object Removal via Tuning-Free Latents Alignment (NeurIPS 2024 · similarity 0.85)
- Paint by Inpaint: Learning to Add Image Objects by Removing Them First (CVPR 2025 · similarity 0.85)
- ROSE: Remove Objects with Side Effects in Videos (NeurIPS 2025 · similarity 0.85)
- GeoComplete: Geometry-Aware Diffusion for Reference-Driven Image Completion (NeurIPS 2025 · similarity 0.85)
- MetaShadow: Object-Centered Shadow Detection, Removal, and Synthesis (CVPR 2025 · similarity 0.84)
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.