KlingAIResearch /ShotStream

About ShotStream

ShotStream is research code that generates multi-shot video frames as you go. It's built on Wan2.1-T2V-1.3B and reports 16 FPS on one NVIDIA GPU. You get inference and training scripts, but it's a reference implementation, not an app. Training needs multi-node GPUs and edits to the bash scripts.

Best for: Researchers and engineers who want to run, fine-tune or study a streaming multi-shot video model.

• Generates frames on the fly for interactive storytelling
• Runs autoregressive 4-step long multi-shot inference
• Reports 16 FPS on a single NVIDIA GPU
• Trains a bidirectional next-shot teacher model
• Initializes the causal student from teacher ODE pairs, following CausVid
• Distills in two stages: intra-shot, then inter-shot self-forcing

1. 1
   Clone the repo and make a conda environment: the command below, the next command, conda activate shotstream

   git clone https://github.com/KlingAIResearch/ShotStream.git

   Copy

   conda create -n shotstream python=3.10 -y

   Copy
2. 2
   Install CUDA, PyTorch and the rest: the command below, then the next command. Or run bash tools/setup/env.sh

   pip install -r requirements.txt

   Copy

   pip install flash-attn --no-build-isolation

   Copy
3. 3
   Download Wan-T2V-1.3B and ShotStream checkpoints into wan_models and ckpts with git-lfs. Or run the command below

   bash tools/setup/download_ckpt.sh

   Copy
4. 4
   Run 4-step long multi-shot generation: the command below

   bash tools/inference/causal_fewsteps.sh

   Copy
5. 5
   To train, set MASTER_ADDR in the bash scripts, then start with the command below

   bash tools/train/1_basemodel.sh 0

   Copy