Latest upload videos examples has the latest workflow embedded into the preview image of the workflow. Its not done but the paste should work better, and simpler cutout options.

Full version should be out in a few days, I am working on all 3 flows, simple, normal and advanced. Let me know if you know anything that is missing you want.

Uploaded wrong version, fixed now.

Advanced

V4(3.5 beta)

First let me say this was a LOT of work and is probably the most powerful workflow i have made. If you like this leave a like, i dont think you will find any flows like this. I tried to make it simple... but im not simple so it may have gotten difficult.

I put every trick i know about videos in here and it has not disapointed me at all, silly thing is i can still think of more to add...

Works for

Text to video

Image to video (first) i cant get last to work alone

Video to video

Ref Image to video

Inpaint video

All of these can be combined or used alone.

Advanced Features

-Inpainting with controls or blank slate

-Cutout inpainting with paste back to original video (beta)

-Extending videos with first frame and ref image. (vid to vid or image to vid)

-Sam1/Sam2/Florence masking options

-watermarks removal

added

-custom video masking

-masking cutout and paste back inpainting (beta)

Fixed

-reactor face paste

Changed

-Everything

-New UI

-Everything is optional

V3 - The holy shit its perfect release.

added

-Added ping pong option

-added first/last frame options making extending work

-text to vid working

- first/last frame working

-image ref only working

-control vid only working

Enable what you want to use.

reworked

-Video processing

-image ref options simplified

removed - Kept for last frame fixing if you want it, its not needed.

- Reactor (extra flow included with it)

https://civarchive.com/models/1645123/video-face-fixing

New loras are out for wan

https://huggingface.co/Kijai/WanVideo_comfy/tree/main

Accvid Lora + causvid = better even less steps, 4 worked for tests

This combo of 0.8 strength on both improved faces with 4 steps compared to 8 with just causvid (you can play with strengths)

Movigen lora - Tested on meme dance, it made the chair dance with the girl...

New causvid loras

V1.5 same as 1.0 but tries to fix flashing at start

v2 needs more steps 8-12? and is helped with small CFG.

Accvid + V2 has not been tested.

TIPS

-Lineart can be generated from a depth map.

-Connect the depth control to lineart and then send that to the control vid

-This provides a lineart outline that works to allow more changes than full lineart.

-this works well with 1.3b

This can inpaint faces with sd1.5 if reactor does not work or you dont use it

https://civarchive.com/models/1116580/sd-15-inpaint-video-enhancing

https://civarchive.com/models/1645123/video-face-fixing

This can add facial expressions if needed

https://civarchive.com/models/1374030/live-portrait-video-editing

This can lip sync you videos

https://civarchive.com/models/1112629/audio-to-video-lipsync-latentsync

This uses WAN 2.1 first/last frames with wan 14b or wan 1.3b INP models

https://civarchive.com/models/1300494/wan-21-video-generation-img2vid-with-fun14b-image-to-video

Advanced

V2 - Full version that i was working on before needing to fix things.

- added full image to video

-- manual sizing and frames for img to vid

To use Image to video Use the switch by the sampler and set manual H/W/FPS/Frames

Includes 2 workflow image examples, one for image to vid set up, and one with vid to vid set up. The pngs can be dragged into comfy to load them with those settings. Imgs and vid refs provided.

NOTE - IMG to vid Switch was not supposed to be included this version but i forgot to remove

-Added image to video switch to disable controls... tested now (seems to work fine, not exactly image to video but close and white bg can allow a prompted bg.)

if using this set your fps manually to the combines to 16

If using this it requires the First frame of a video to get the size, i did not set up custom sizing yet so its buggy and uses the first frame of a video to size your generation even if its not using the controls.

V1.1

Fixed

-removed problematic nodes and replaced with node sets in the flow already or native.

-added switch bypass to white bg to allow passthrough of original image background

Added

-.webp loading

-Workflow save at end to save PNG image of the workflow and settings

-last frame image (i dont know if it can be used as ref for a second batch)

-Replaced imagetobatch+ with impact node

-Added dual subject ref creation and use

-Added fixed BG remove/replace for both single and dual.

-changed repeat frames to process videos once and repeat controls

-separated sizing for render and sizing for controls to prevent needing to redo controls to change render size.

-moved sizing to near sampler

-no longer resizing twice.

--reactor changes (ignore if you remove)

-Added dual person reactor face fix (still not sure if this works properly)

-Added reactor mask helper to fix hair distortions after using it.

Ultra simple

Was asked to make a flow using a little as possible and still work.

This has everything removed but still works, just does not offer same quality or options.

I have taken this time to try to remove some problematic nodes that some have had issues installing. Let me know if this has any of those errors.

Things kept

Single image ref on white

depth/pose/lineart controls

repeating looped videos

Model upscale

Frame interpolation

V2 - simplified

Hotfix to fix video crop error.

Video ref pack is a bunch of video refs that should work with the flow.

Enjoy :)

Vace vid to vid

V1

Video to video with Vace ref image.

Hotfixed the resize to ensure sizing, divisible by 16

Models and loras

https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/tree/main/split_files/diffusion_models

GGUF

https://huggingface.co/QuantStack/Wan2.1-VACE-14B-GGUF/tree/main

Vace comfy info

https://docs.comfy.org/tutorials/video/wan/vace

Vace 8 step loras

https://huggingface.co/Kijai/WanVideo_comfy/tree/main

I only tested the lora with gguf.