| | --- |
| | license: cc-by-nc-4.0 |
| | base_model: zibojia/minimax-remover |
| | tags: |
| | - video |
| | - inpainting |
| | - object-removal |
| | - diffusers |
| | - safetensors |
| | --- |
| | |
| | # MiniMax-Remover — Video Object Removal |
| |
|
| | Mirror of [zibojia/minimax-remover](https://huggingface.co/zibojia/minimax-remover) for use with [ComfyUI-FFMPEGA](https://github.com/AEmotionStudio/ComfyUI-FFMPEGA). |
| |
|
| | ## What is MiniMax-Remover? |
| |
|
| | MiniMax-Remover is a purpose-built DiT (Diffusion Transformer) model for **video object removal**. Given a video and a binary mask indicating unwanted regions, the model inpaints the masked areas with temporally consistent content. |
| |
|
| | Key features: |
| | - **81-frame native batch size** — processes up to 81 frames at once for temporal consistency |
| | - **6–12 inference steps** — fast inference with iterative mask dilation |
| | - **Simplified DiT architecture** — lightweight compared to general-purpose video editors (~2.5 GB) |
| |
|
| | ## Files |
| |
|
| | ``` |
| | transformer/diffusion_pytorch_model.safetensors (~2.25 GB) |
| | vae/diffusion_pytorch_model.safetensors (~508 MB) |
| | scheduler/scheduler_config.json |
| | ``` |
| |
|
| | ## Usage |
| |
|
| | ### With ComfyUI-FFMPEGA (recommended) |
| |
|
| | 1. Enable the `use_minimax_remover` toggle on the FFMPEG Agent node |
| | 2. Use `auto_mask:effect=remove` or select `minimax_remover` no-LLM mode |
| | 3. The model auto-downloads on first use |
| |
|
| | ### Manual download |
| |
|
| | ```bash |
| | huggingface-cli download AEmotionStudio/minimax-remover --local-dir ./minimax_remover |
| | ``` |
| |
|
| | ### Programmatic |
| |
|
| | ```python |
| | from huggingface_hub import snapshot_download |
| | |
| | snapshot_download( |
| | repo_id="AEmotionStudio/minimax-remover", |
| | allow_patterns=["vae/*", "transformer/*", "scheduler/*"], |
| | local_dir="./minimax_remover" |
| | ) |
| | ``` |
| |
|
| | ## Removal Priority in FFMPEGA |
| |
|
| | When used with ComfyUI-FFMPEGA, MiniMax-Remover has the **highest priority** for object removal: |
| |
|
| | 1. **MiniMax-Remover** (~2.5 GB VRAM) — when `use_minimax_remover=On` |
| | 2. **FLUX Klein 4B** (~15 GB VRAM) — when `use_flux_klein=On` |
| | 3. **LaMa** (~200 MB VRAM) — always available fallback |
| | 4. **Black fill** (0 VRAM) — FFmpeg fallback |
| |
|
| | ## ⚠️ License |
| |
|
| | > **Non-Commercial Use Only**: Model weights are licensed under **CC-BY-NC-4.0**. The source code is Apache 2.0. |
| | > |
| | > Users must accept the non-commercial license terms when downloading. Commercial use of the model weights requires separate licensing from the original authors. |
| |
|
| | ## Credits |
| |
|
| | - Original model by: [Jiaqi Zhu et al.](https://github.com/zibojia/MiniMax-Remover) |
| | - Paper: *"MiniMax-Remover: Taming Bad Noise Helps Video Object Removal"* |
| | - Upstream HuggingFace: [zibojia/minimax-remover](https://huggingface.co/zibojia/minimax-remover) |
| | - Redistributed by: [Æmotion Studio](https://huggingface.co/AEmotionStudio) for use with [ComfyUI-FFMPEGA](https://github.com/AEmotionStudio/ComfyUI-FFMPEGA) |
| |
|
