# From Draft to Draft-Free: One-Step Video Object Removal via Privileged Distillation and Fast Planting **Zizhao Chen, Ping Wei†, Guang Dai, Jingdong Wang, Mengmeng Wang†** Xi'an Jiaotong University  ·  SGIT AI Lab  ·  Zhejiang University of Technology  ·  Baidu † Corresponding authors Project Page arXiv Paper Hugging Face GitHub Code

Overview of the D2DF framework

**D2DF** is a three-stage framework for one-step video object removal. It first distills a multi-step draft-guided diffusion teacher into **D2DF-DG**, a one-step draft refiner, through **Prior-Privileged Consistency Distillation (PPCD)**. It then uses the lightweight **Self-Guided Fast Planting (SGFP)** module to construct pseudo-drafts in latent space, producing **D2DF-DF**, a fully draft-free one-step generator. This repository provides standalone inference code for both D2DF-DF and D2DF-DG using the official Hugging Face `diffusers` and `transformers` packages. ## Installation ```bash git clone https://github.com/bigD233/D2DF.git cd D2DF pip install -r requirements.txt ``` We recommend Python 3.11 and an NVIDIA GPU with CUDA support. The released inference code was tested with PyTorch 2.8, CUDA 12.8, `diffusers >= 0.32`, and `transformers >= 4.44`. ## Model Weights Download the base **CogVideoX-5B-I2V** model from its [official Hugging Face repository](https://huggingface.co/zai-org/CogVideoX-5b-I2V). Download the **D2DF-DF** and **D2DF-DG** checkpoints from our [D2DF Hugging Face repository](https://huggingface.co/BigD233333/D2DF). After extracting the downloaded archives, pass the corresponding directories to `--transformer_path`. The included `prompt_embeds.pt` stores the pre-computed embedding for the prompt `"background"`, so the T5 encoder does not need to be loaded during normal inference. ## Input Format Prepare the input as directories of numbered image files: ```text inputs/ ├── imgs/ │ ├── 00000.jpg │ ├── 00001.jpg │ └── ... ├── mask/ │ ├── 00000.png │ ├── 00001.png │ └── ... └── drafts/ # required only by D2DF-DG ├── 00000.jpg ├── 00001.jpg └── ... ``` - The repository includes this ready-to-run 25-frame example under `inputs/`. - Frames, masks, and drafts are read in filename order. - Mask pixels belonging to the removal target should have value `255` by default. - Inputs are resized to the requested output resolution. - The default setting uses 25 frames at 480 x 720 resolution. ## Inference ### D2DF-DF: Draft-Free One-Step Removal No external draft is required: ```bash python infer.py \ --mode df \ --base_model_path \ --transformer_path \ --frames_dir inputs/imgs \ --mask_dir inputs/mask \ --output_dir outputs/I-210910_I09032_T04_df \ --num_frames 25 \ --height 480 \ --width 720 ``` ### D2DF-DG: Draft-Guided One-Step Refinement Supply a draft produced by a video object removal method such as ProPainter: ```bash python infer.py \ --mode dg \ --base_model_path \ --transformer_path \ --frames_dir inputs/imgs \ --mask_dir inputs/mask \ --draft_dir inputs/drafts \ --output_dir outputs/I-210910_I09032_T04_dg \ --num_frames 25 \ --height 480 \ --width 720 ``` The default configuration performs one denoising step with `guidance_scale=1.0`. The output directory contains an MP4 video and a GIF preview. Add `--save_frames` to also save individual PNG frames. ## Citation ```bibtex @misc{chen2026d2df, title={From Draft to Draft-Free: One-Step Video Object Removal via Privileged Distillation and Fast Planting}, author={Zizhao Chen and Ping Wei and Guang Dai and Jingdong Wang and Mengmeng Wang}, year={2026} } ``` ## Acknowledgements Our implementation is built upon [CogVideoX](https://github.com/THUDM/CogVideo), [Hugging Face Diffusers](https://github.com/huggingface/diffusers), and [ProPainter](https://github.com/sczhou/ProPainter). We sincerely thank the contributors of these projects for their excellent work. ## License This repository is released under the [Apache License 2.0](LICENSE). ## Contact For questions about this work, please contact [Zizhao Chen](mailto:bigd_chen@stu.xjtu.edu.cn).