LongCat-Video Examples¶

These examples provide text-to-video, image-to-video, video continuation, unified generation, and refinement with LongCat-Video.

Model Source¶

Model	HuggingFace	ModelScope	Purpose
LongCat-Video	meituan-longcat/LongCat-Video	meituan-longcat/LongCat-Video	Base video generation
Wan2.1 VAE	Wan-AI/Wan2.1-T2V-14B	Wan-AI/Wan2.1-T2V-14B	Video encoder and decoder
RIFE v4.26	RIFEv4.26	N/A	Optional frame interpolation

Feature Support¶

Feature	Support	Notes
Text-to-video	Supported	Standard and refinement entry points
Image-to-video	Supported	Requires a source image
Video continuation	Supported	Requires a source video
Multi-GPU inference	Supported	CFG and Ulysses parallelism through `--gpu_num`
LoRA	Supported	Distillation and refinement LoRAs
Quantization	Unsupported	No quantized example is checked in
CPU offload	Supported	Base scripts use model offload for memory control
Feature cache	N/A	The implementation uses its model-specific KV cache
Server API	Supported	Generation examples expose standard pipeline functions

Requirements¶

GPU: one CUDA GPU for base examples; additional GPUs may be selected through --gpu_num
Software: the standard TeleFuser installation; RIFE is required only for frame interpolation
Input assets: a readable image for I2V or a readable video for continuation

Install TeleFuser by following the development setup.

Model Directory¶

${TF_MODEL_ZOO_PATH}/
|-- LongCat-Video/
|   |-- diffusion_pytorch_model-*.safetensors
|   |-- text_encoder/
|   |-- tokenizer/
|   \-- lora/
|       \-- refinement_lora.safetensors
|-- Wan2.1-T2V-1.3B/
|   \-- Wan2.1_VAE.pth
\-- RIFEv4.26_0921/
    \-- flownet.pkl

export TF_MODEL_ZOO_PATH=/path/to/model_zoo

Quick Start¶

TELEAI_EXAMPLE_OUTPUT_DIR=work_dirs \
python examples/longcat_video/longcat_text_to_video.py \
  --model_root "$TF_MODEL_ZOO_PATH/LongCat-Video" \
  --prompt "A boat sailing across a calm ocean"

The command writes work_dirs/longcat_text_to_video_1gpu.mp4.

Examples¶

Text-To-Video¶

`longcat_text_to_video.py`¶

python examples/longcat_video/longcat_text_to_video.py \
  --model_root "$TF_MODEL_ZOO_PATH/LongCat-Video" \
  --resolution 720p --aspect_ratio 16:9 \
  --prompt "A wide landscape with slow camera motion"

Image-To-Video¶

`longcat_image_to_video.py`¶

python examples/longcat_video/longcat_image_to_video.py \
  --model_root "$TF_MODEL_ZOO_PATH/LongCat-Video" \
  --image_path /path/to/input.jpg \
  --prompt "Natural motion in the scene"

Video Continuation¶

`longcat_video_continue.py`¶

python examples/longcat_video/longcat_video_continue.py \
  --model_root "$TF_MODEL_ZOO_PATH/LongCat-Video" \
  --video_path /path/to/input.mp4 \
  --prompt "Continue the scene naturally"

Unified Tasks¶

`longcat_video_unify.py`¶

Use the unified pipeline when one loaded model must serve multiple generation modes:

python examples/longcat_video/longcat_video_unify.py \
  --model_root "$TF_MODEL_ZOO_PATH/LongCat-Video" \
  --task t2v --prompt "A sunset over the sea"

`longcat_text_to_video_refine.py`¶

python examples/longcat_video/longcat_text_to_video_refine.py \
  --model_root "$TF_MODEL_ZOO_PATH/LongCat-Video" \
  --height 480 --width 832 \
  --refine_height 720 --refine_width 1280

Key behavior:

Base generation supports one or more GPUs through --gpu_num.
Refinement can enable block-sparse attention and temporal extension.