Tiny Wan2.1 VSA T2V 14B 720P Debug Pipeline

This is a randomly initialized, tiny WanPipeline fixture for FastVideo/Wan2.1-VSA-T2V-14B-720P-Diffusers. FastVideo Wan2.1 VSA 14B 720P text-to-video layout represented as a tiny Diffusers WanPipeline artifact for FastVideo VideoGenerator and VIDEO_SPARSE_ATTN load-path debugging. The saved transformer weights include FastVideo VSA to_gate_compress tensors, which are not emitted by Diffusers.

It is intended for fast load-path and inference-control debugging only. It is not trained and should not be used for generation quality evaluation.

import os
from fastvideo import VideoGenerator

os.environ["FASTVIDEO_ATTENTION_BACKEND"] = "VIDEO_SPARSE_ATTN"
generator = VideoGenerator.from_pretrained(
    "Erland/tiny-wan2.1-vsa-t2v-14b-720p-debug",
    num_gpus=1,
    pipeline_config={"flow_shift": 5.0},
    VSA_sparsity=0.5,
)
try:
    generator.generate_video(
        "debug prompt",
        output_path="my_videos/",
        save_video=True,
        height=64,
        width=64,
        num_frames=5,
        num_inference_steps=3,
        guidance_scale=1.0,
    )
finally:
    generator.shutdown()
Downloads last month
21
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support