Inpaint Anything
github.com
- Windows
- Web
- Self-hosted
What Inpaint Anything is
Inpaint Anything is a free, open-source project for removing, filling, or replacing objects in images, videos, and 3D scenes. Users select an object for segmentation, then an inpainting model such as LaMa fills the resulting area. Fill Anything generates content in a selected area from a text prompt, while Replace Anything uses a prompt to change a background. For video, the workflow tracks an object across frames and fills the affected areas with a video inpainting model such as STTN. For 3D scenes, it tracks an object across source views, inpaints those views, and uses NeRF to synthesize new views without the object. The project describes support for any aspect ratio and 2K resolution. Users can run a local web interface on their own machine or try a hosted interface linked from the README. Installation requires Python 3.8 or newer, PyTorch, model-specific dependencies, and downloaded model checkpoints. The Windows setup recommends Miniconda and an administrator Anaconda PowerShell Prompt. A separate beta branch has newer requirements and targets a single 24GB GPU by default.
Who it suits
Inpaint Anything suits people who want to remove or modify objects across images, video, or 3D scenes and are comfortable installing dependencies and model checkpoints. Users considering commercial work should review the stated model-license restrictions.
Good to know, and what to watch
- Works with images, videos, and 3D scenes.
- Text prompts guide fill and background replacement.
- Provides a local web interface.
- Describes support for any aspect ratio and 2K resolution.
- Installation requires Python, PyTorch, and model-specific dependencies.
- Model checkpoints must be downloaded and placed in project directories.
- Default hardware target is a single 24GB GPU.
- Some models have non-commercial licenses.
The verdict
Inpaint Anything offers several object-editing workflows and can run locally or through a linked hosted interface. Installation requirements, hardware needs, and model licenses are practical considerations before use.
What the maker publishes
- 3D processing
- Remove Anything 3D tracks a selected object across source views, inpaints those views, and uses NeRF to synthesize novel views without the object. source
- Beta branch
- The README describes the main_2026 branch as beta and says it requires Python 3.12 or newer, PyTorch 2.7 or newer, and CUDA 12.6 or newer. source
- Fill and replace
- Text prompts guide filling selected objects with new content or replacing backgrounds. source
- Founded
- 2023 source
- Hardware
- The repository says its defaults target a single 24GB GPU, and FLUX.1-Fill can require about 20GB of VRAM. source
- Hosted demo
- The README links to a Hugging Face interface for trying Inpaint Anything online. source
- Image capabilities
- The README lists support for any aspect ratio and 2K resolution. source
- Integrations
- The README identifies Segment Anything, LaMa, Stable Diffusion, OSTrack, and STTN as models used or acknowledged by the project. source
- License
- The repository includes an Apache License, Version 2.0, which grants no-charge copyright and patent permissions subject to its terms. source
- Local interface
- The project provides a local web UI that users can run on their own machine. source
- Local interface
- The repository says its demo can run as a local web UI; the main_2026 quick start launches it at localhost:7860. source
- Media support
- The repository describes workflows for images, videos, and 3D scenes, with support for any aspect ratio and 2K resolution. source
- Model downloads
- Users need to download model checkpoints and place them in the documented project directories to run the workflows. source
- Model licensing
- The README notes that ProPainter and FLUX.1-Fill-dev are non-commercially licensed and says to check the terms for LaMa, SDXL, and SAM 3 before commercial use. source
- Models
- The main_2026 branch uses SAM 3, LaMa, SDXL, and ProPainter by default in different stages, with other backends available. source
- Object removal
- Users click an object for segmentation, then an inpainting model such as LaMa fills the resulting hole. source
- Object removal
- Users can select an object by clicking it or, on the main_2026 branch, naming it with a text prompt, then fill the removed area. source
- Prompted editing
- Fill Anything uses a text prompt to generate content in a selected area, while Replace Anything uses a prompt to change the background. source
- Purpose
- Inpaint Anything uses Segment Anything and inpainting models to remove, fill, or replace objects in images, videos, and 3D scenes. source
- Purpose
- Inpaint Anything uses segmentation and inpainting models to edit images, videos, and 3D scenes. source
- Requirements
- The documented installation requires Python 3.8 or newer and lists PyTorch and model-specific dependencies. source
- Requirements
- The main_2026 branch requires Python 3.12 or later, PyTorch 2.7 or later, and CUDA 12.6 or later for its default SAM 3 setup. source
- Research authors
- The project lists Tao Yu, Runseng Feng, Ruoyu Feng, Jinming Liu, Xin Jin, Wenjun Zeng, and Zhibo Chen as authors. source
- Robotics workflow
- The main_2026 branch includes batch hand removal from egocentric footage and mask export for Human-to-Robot pipelines. source
Checked against the maker's pages on 9 Oct 2026.