GeoRemover: Removing Objects and Their Causal Visual Artifacts
Paper • 2509.18538 • Published • 1
import torch
from diffusers import DiffusionPipeline
from diffusers.utils import load_image
# switch to "mps" for apple devices
pipe = DiffusionPipeline.from_pretrained("buxiangzhiren/GeoRemover", dtype=torch.bfloat16, device_map="cuda")
prompt = "Turn this cat into a dog"
input_image = load_image("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/cat.png")
image = pipe(image=input_image, prompt=prompt).images[0]GeoRemover: Removing Objects and Their Causal Visual Artifacts (NeurIPS 2025 • Spotlight)
Paper: GeoRemover: Removing Objects and Their Causal Visual Artifacts — https://arxiv.org/abs/2509.18538
Code: https://github.com/buxiangzhiren/GeoRemover
Authors: Zixin Zhu, Haoxiang Li, Xuelu Feng, He Wu, Chunming Qiao, Junsong Yuan
Abstract (short)
Object removal should delete both the target object and its causal visual artifacts (e.g., shadows, reflections). We propose a geometry-aware two-stage framework: (1) geometry removal on depth with strict mask-aligned supervision; (2) appearance rendering to RGB conditioned on the updated geometry. A preference-driven objective guides stage-1 to remove objects and their artifacts while avoiding new structural insertions.
Model description
- Two stages:
- Stage-1: Geometry Removal — edit depth/geometry under strict mask supervision to remove objects at the structural level.
- Stage-2: Appearance Rendering — render photorealistic RGB conditioned on the edited depth so shadows/reflections adapt consistently.
- Implementation is built on black-forest-labs/FLUX.1-Fill-dev with LoRA adapters for both stages.
- This repo hosts the LoRA weights (stage-1 and/or stage-2). Depth maps are obtained with Video-Depth-Anything v2.
Intended uses
- Research on object removal where artifact consistency (shadows/reflections) matters.
- Scenes where monocular depth is reasonably reliable.
Limitations
- Quality depends on depth accuracy; glass/transparent objects, extreme lighting, or heavy blur may fail.
- Trained primarily on RORD; cross-domain generalization may vary.
- Non-commercial research use (see License). Comply with the base model’s license.
- Downloads last month
- -