Skip to content

first_generation.md

Latest commit

 

History

History
46 lines (31 loc) · 2.86 KB

first_generation.md

File metadata and controls

46 lines (31 loc) · 2.86 KB
Raw

Generate your first video

This walkthrough uses Wan2.1-T2V-1.3B. Complete the environment setup first, then run the commands below from the LightX2V repository root.

Download the model

hf download Wan-AI/Wan2.1-T2V-1.3B --local-dir /path/to/models/Wan2.1-T2V-1.3B

Alternatively, download Wan-AI/Wan2.1-T2V-1.3B through ModelScope and use its local directory as model_path.

Configure and run

Open run_wan_t2v.sh and set:

Setting Value
lightx2v_path Absolute path to this repository
model_path Absolute path to the downloaded model
CUDA_VISIBLE_DEVICES GPU to use, such as 0
--prompt / --negative_prompt Text describing the scene and unwanted content
--size Output size as HEIGHT WIDTH; the launcher uses 480 832
--num_frames Output frame count; the launcher uses 81
--save_result_path Output video path

The launcher selects wan_t2v.json. Edit startup settings such as sampling steps and guidance scale in JSON, and request parameters in the launcher.

The launcher's explicit --size and --num_frames override the corresponding JSON defaults, so changing only JSON does not change this launcher's output size or frame count. Edit those arguments in the launcher. To control both through JSON instead, remove the two arguments from the launcher, then set size ([height, width]) and num_frames in JSON.

The supplied JSON uses flash_attn3. Select attention operators supported by your GPU and installation; see the attention guide. For lower GPU memory usage, see the offload guide.

bash scripts/wan/run_wan_t2v.sh

Wait for inference to finish and open the file at --save_result_path. The supplied launcher writes to save_results/output_lightx2v_wan_t2v.mp4 under the repository.

Continue

  • Wan2.1 and VACE: model variants, distilled weights, services and multiple GPUs.
  • Wan2.2 and Qwen Image: model-specific launchers and configurations.
  • Python API: embed generation in an application and reuse a pipeline.
  • Service deployment: serve generation requests. The Wan-specific server and client are listed in the Wan README.

Use the same model, JSON and request parameters, including seed, size and frame count, when comparing CLI, service and Python execution. The manually configured Python examples may intentionally use different sampling settings.