コンパクトなモデルで、数分で生成できる
A100 GPUで数分で生成。A10でも生成できる。しかもそこそこ綺麗。
動画生成AIの中ではトップレベルでコスパいい。
いろんな制御方法があるので、例は随時追記していきます。
Image2Video
python vace/vace_preproccess.py --task image_reference \
--mode salient \
--image "girl.jpg"
python vace/vace_wan_inference.py \
--src_ref_images "processed/image_reference/{DATE}/src_ref_image-image_reference.png" \
--prompt "A girl is smiling and walking slowly towards the camera. Highly detailed facial drawing."
Face2Video
python vace/vace_preproccess.py --task image_face --image processed/image_reference/2025-06-01-01-13-38/src_ref_image-image_reference.png
python vace/vace_wan_inference.py \
--ckpt_dir models/Wan2.1-VACE-1.3B/ \
--src_ref_images processed/image_face/{DATE}/src_ref_image-image_face.png \
--prompt "Girl in robot suit shooting lasers from her hands. Highly detailed facial features."
Video2Depth2Video
python vace/vace_preproccess.py --task image_face --image processed/image_reference/2025-06-01-01-13-38/src_ref_image-image_reference.png
python vace/vace_wan_inference.py \
--ckpt_dir models/Wan2.1-VACE-1.3B/ \
--src_ref_images processed/image_face/{DATE}/src_ref_image-image_face.png \
--prompt "Girl in robot suit shooting lasers from her hands. Highly detailed facial features."
VideoInpaint2Video
元動画 / 前処理結果 / 出力動画

python vace/vace_preproccess.py --task inpainting --mode salientmasktrack \
--video 12744585_1080_1920_100fps.mp4 --label person --maskaug_mode original_expand --maskaug_ratio 0.3
python vace/vace_wan_inference.py \
--src_video processed/inpainting/{DATE}/src_video-inpainting.mp4\
--src_mask processed/inpainting/{DATE}/src_mask-inpainting.mp4 \
--prompt "老虎望着大海,站起身走开。场景逼真。"