For an introduction and the list of main configuration/file setup parameters to get you started, refer to the earlier section at: https://gist.github.com/Glidias/b15b51598ae643bab9dbc7aa12fe62ed
These doc only covers anything workflow specific beyond file/json-based stuff.
This is normally used only when inpainting localised regions across multiple (or single-image) frames. Can disable this if not needed.
There is a group to save a variant of the output video WEBP as an inpaint composite consisting of the output video frames edge-blended in with the originally set up VACE source control video frames through VACE control video mask (if any) to help preserve the original quality of the source image areas that aren't inpainted.
How it works? The edges of the inverted control mask is expanded out and blurred by half of the expansion amount for the blur radius to ensure the output is fully and seamlessly blended in over the original control video images. The blur radius varies depending on the resolution of the video (higher resolution means larger blur radius), but is based off similar to the Model Shift formula conventions.
WEBP image saved in this group are suffixed with --blend
, so you get {stem}--stem
as the filename prefix.
In most cases, this output variant should not yield any artifacts with the blended version for videos where every frame is inpainted with a localised region for masking and there are no brand new frames generated, thus this allows you to regain the original quality of various portions of the original source image frames you had initially provided for inpainting. But of course, if your output video's saturation differs greatly from the originally provided source images, the results might not work well even with blending at the edges.
This section covers how the conventions are adapted for other use cases in none-VACE workflows (eg. WAN I2V/FLV2V/etc.)
Calculates total frames to use from respective json/psd/video source definition or else uses the workflow default preset.
Assumes the source assets have no unwanted transparent regions in alpha channel (if any) (as these will become black and there are no VACE mask inpainting options available).
Any mask/extramask assets are ignored entirely.
For this workflow case under both I2V and FLF2V, the resulting "first" and "last" source images can be modified via the V2V Frame length control custom JSON parameters settings for "load_cap":
and "skip_index":
respectively, if needed be.
For folder of images, will treat the last source frame image (alphabetical order) to use as starting image for I2V. For video, will treat the last source frame image (alphabetical order) to use as starting image for I2V. For regular image, will use the image itself.
Variants: FusionX Ingredients, Standard WAN 2.1/2.2, etc.
Will all source assets, will treat the first source frame as first frame and last source frame as last frame. Any source frames in between are ignored.
Variants: Standard WAN 2.1/2.2, etc.