SolitaryThinker
825c6365e0
A bundle's training variant (header-declared `variant`, filename token as
fallback) now selects its sampling preset: distilled -> ltx2_distilled
(8 steps, cfg 1.0, distilled sigmas engage at <=8 steps), sft/base ->
ltx2_base. The bundle table also pins the pipeline class, so
VideoGenerator.from_pretrained(<bundle>) resolves without an override.
Adds basic_ltx2_5{,_distilled}.py examples (--model-path takes the FILE,
--gemma-root declares the paired encoder root), a README section, and CPU
tests for variant/preset resolution, table integrity, and the examples.
Basic Video Generation Tutorial
The VideoGenerator class provides the primary Python interface for doing offline video generation, which is interacting with a diffusion pipeline without using a separate inference api server.
Requirements
- At least a single NVIDIA GPU with CUDA 12.4.
- Python 3.10-3.12
Installation
If you have not installed FastVideo, please following these instructions first.
Usage
The first script in this example shows the most basic usage of FastVideo. If you are new to Python and FastVideo, you should start here.
# if you have not cloned the directory:
git clone https://github.com/hao-ai-lab/FastVideo.git && cd FastVideo
python examples/inference/basic/basic.py
For an example on Apple silicon:
python examples/inference/basic/basic_mps.py
For an example running DMD+VSA inference:
python examples/inference/basic/basic_dmd.py
For the typed config/request path added during the inference API refactor:
python examples/inference/basic/basic_dmd_new_api.py
LTX-2.5 single-file bundles
These examples load LTX-2.5 from a single-file bundle: one .safetensors
file carrying every component (transformer, video VAE, audio VAE, vocoder,
text projection). (The Hugging Face release ships as a split pack — one
file per component — which loads through the standard repo path instead.)
Pass the bundle FILE path directly:
python examples/inference/basic/basic_ltx2_5_distilled.py \
--model-path /path/to/bundle.safetensors \
--gemma-root /path/to/gemma_root
| bundle variant | preset | sampling defaults |
|---|---|---|
| distilled | ltx2_distilled |
one stage, 8 steps, CFG 1.0 (no CFG/STG, no spatial upscaler) |
| sft / base | ltx2_base |
40 steps, CFG 3.0 with STG |
- The variant is read from the bundle header when declared; otherwise the
distilledtoken in the file name decides. - The Gemma text encoder is NOT in the bundle. Declare its root with
--gemma-root(orFASTVIDEO_LTX_ENCODER_ROOT), and use the root shipped with the same transformer variant — the roots differ in prompt templating. - Decoder caveat: bundles that declare a diffusion VAE decoder
(
CausalDiffusionVAE) are not supported yet; loading them currently fails at VAE build time. Bundles with the classicCausalVideoAutoencoderrun end to end.
Basic Walkthrough
All you need to generate videos using multi-gpus from state-of-the-art diffusion pipelines is the following few lines!
from fastvideo import VideoGenerator
def main():
generator = VideoGenerator.from_pretrained(
"Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
num_gpus=1,
)
prompt = ("A curious raccoon peers through a vibrant field of yellow sunflowers, its eyes "
"wide with interest. The playful yet serene atmosphere is complemented by soft "
"natural light filtering through the petals. Mid-shot, warm and cheerful tones.")
video = generator.generate_video(prompt)
if __name__ == "__main__":
main()