Vivid-VR: Distilling Concepts from Text-to-Video Diffusion Transformer for Photorealistic Video Restoration

Haoran Bai,  Xiaoxu Chen,  Canqian Yang,  Zongyao He,  Sibin Deng,  Ying Chen^∗

Alibaba Group - Taobao & Tmall Group

* Corresponding author

Paper | Project Page

Click to download more visual comparisons on [synthetic], [real-world], and [AIGC] videos.
For more video visualizations, go checkout our [project page].

---

🔥 Update

• [2026.01.26] This work has been accepted by ICLR 2026.
• [2025.09.29] Paper Update: Add some experiments. Go checkout the revised paper at [link].
• [2025.09.20] Support restoration-guided sampling for the trade-off between fidelity and realism, using the arg "--restoration_guidance_scale", click here to view visual comparisons.
• [2025.08.30] Support long video inference by aggregate sampling in the temporal dimension, using the arg "--num_temporal_process_frames".
• [2025.08.21] Paper is released at [link].
• [2025.08.06] UGC50 and AIGC50 testsets are made publicly available from [link].
• [2025.08.05] Inference code is released.
• [2025.08.05] This repo is created.

🎬 Overview

🔧 Dependencies and Installation

1. Clone Repo

   git clone https://github.com/csbhr/Vivid-VR.git
   cd Vivid-VR

2. Create Conda Environment and Install Dependencies

   # create new conda env
   conda create -n Vivid-VR python=3.10
   conda activate Vivid-VR
   
   # install pytorch
   pip install torch==2.2.1 torchvision==0.17.1 torchaudio==2.2.1 --index-url https://download.pytorch.org/whl/cu121
   
   # install python dependencies
   pip install -r requirements.txt
   
   # install easyocr [Optional, for text fix]
   pip install easyocr
   pip install numpy==1.26.4  # numpy2.x maybe installed when installing easyocr, which will cause conflicts

3. Download Models

   • [Required] Download CogVideoX1.5-5B checkpoints from [huggingface].
   • [Required] Download cogvlm2-llama3-caption checkpoints from [huggingface].
     • Please replace modeling_cogvlm.py in the downloaded cogvlm2-llama3-caption directory with ./VRDiT/cogvlm2-llama3-caption /modeling_cogvlm.py to remove the dependency on pytorchvideo.
   • [Required] Download Vivid-VR checkpoints from [huggingface].
   • [Optional, for text fix] Download easyocr checkpoints [english_g2] [zh_sim_g2] [craft_mlt_25k].
   • [Optional, for text fix] Download Real-ESRGAN checkpoints [RealESRGAN_x2plus].
   • Put them under the ./ckpts folder.

   The ckpts directory structure should be arranged as:

   ├── ckpts
   │   ├── CogVideoX1.5-5B
   │   │   ├── ...
   │   ├── cogvlm2-llama3-caption
   │   │   ├── ...
   │   ├── Vivid-VR
   │   │   ├── controlnet
   │   │       ├── config.json
   │   │       ├── diffusion_pytorch_model.safetensors
   │   │   ├── connectors.pt
   │   │   ├── control_feat_proj.pt
   │   │   ├── control_patch_embed.pt
   │   ├── easyocr
   │   │   ├── craft_mlt_25k.pth
   │   │   ├── english_g2.pth
   │   │   ├── zh_sim_g2.pth
   │   ├── RealESRGAN
   │   │   ├── RealESRGAN_x2plus.pth
   

☕️ Quick Inference

Run the following commands to try it out:

python VRDiT/inference.py \
    --ckpt_dir=./ckpts \
    --cogvideox_ckpt_path=./ckpts/CogVideoX1.5-5B \
    --cogvlm2_ckpt_path=./ckpts/cogvlm2-llama3-caption \
    --input_dir=/dir/to/input/videos \
    --output_dir=/dir/to/output/videos \
    --num_temporal_process_frames=121 \  # For long video inference, if video longer than num_temporal_process_frames, aggregate sampling will be enabled in the temporal dimension
    --restoration_guidance_scale=-1.0 \  # Optional, for restoration-guided sampling, if set to -1.0， it will be disable
    --upscale=0 \  # Optional, if set to 0, the short-size of output videos will be 1024
    --textfix \  # Optional, if given, the text region will be replaced by the output of Real-ESRGAN
    --save_images  # Optional, if given, the video frames will be saved

GPU memory usage:

• For a 121-frame video, it requires approximately 43GB GPU memory.
• If you want to reduce GPU memory usage, replace "pipe.enable_model_cpu_offload" with "pipe.enable_sequential_cpu_offload" in ./VRDiT/inference.py. GPU memory usage is reduced to 25GB, but the inference time is longer.
• For the arg "--num_temporal_process_frames", smaller values ​​require less GPU memory but increase inference time.

Trade-off between fidelity and realism:

• Using the arg "--restoration_guidance_scale" to enable restoration-guided sampling. Higher value yield more realistic results, while lower value preserve greater fidelity to the original input content. When the value is -1, restoration-guided sampling is disabled.
• Click here to view visual comparisons.

📧 Citation

If you find our repo useful for your research, please consider citing it:

@InProceedings{bai2026vividvr,
   title={Vivid-VR: Distilling Concepts from Text-to-Video Diffusion Transformer for Photorealistic Video Restoration}, 
   author={Bai, Haoran and Chen, Xiaoxu and Yang, Canqian and He, Zongyao and Deng, Sibin and Chen, Ying},
   booktitle={International Conference on Learning Representations (ICLR)},
   year={2026}
 }

📄 License

• This repo is built based on diffusers v0.31.0, which is distributed under the terms of the Apache License 2.0.
• CogVideoX1.5-5B models are distributed under the terms of the CogVideoX License.
• cogvlm2-llama3-caption models are distributed under the terms of the CogVLM2 License and LLAMA3 License.
• Real-ESRGAN models are distributed under the terms of the BSD 3-Clause License.
• easyocr models are distributed under the terms of the JAIDED.AI Terms and Conditions.