Wan AiO. +VACE14b,Phantom

Optimisation

To use torch.compile and sageattention, run those commands:

pip install triton-windows

pip install sageattention

You need to have MSVC tools and CUDA 12.6+ installed and set in PATH.

For portable also grab include_libs.zip from https://github.com/woct0rdho/triton-windows/releases/tag/v3.0.0-windows.post1 appropriate for your python version and unzip it into python_embeded folder.

To gain free easy speed bost, update comfyui to 0.3.30, pytorch to 2.7.0, if you didn't already, using this command

pip install -U torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu128

then in your run.bat file add '--fast fp16_accumulation' parameter.

Fresh Comfy installs are shipped with this torch version, but you need to add parameter manually, and portable has run_nvidia_gpu_fast_fp16_accumulation.bat.

With both of those and Teacache with moderate settings generation time reduced by 65-70%, without TeCache - by 50-55%.

First launch

If you installed all dependencies and still get red sea of group nodes, first try to close workflow, close tab, restart comfy, open workflow again, then nodes should be detected. If issue persists, there is some conflict in nodepack versions, do this:

You have to make and rollback changes to apply preconfigured values - especially critical for HY nodes where default cfg=8, which will break it. Gotta do it with every node you're planning to use.

If you're wondering, why it can't be done in less frustrating way, remember - comfy is a lie.

A collection of workflows turned into group nodes to be used in a single space - to avoid constant jumping between workflows, saving/uploading videos and swapping values.

A workplace, rather than a workflow.

Generate and extend videos with different lengths and loras indefinitely, create loops, upscale.

Workflow seems massive and complex, until you realize you only have to use few nodes.

Use it, prune it, break it into pieces or just learn it's concepts.

What's in

AiO:

• basic functions - T2V, I2V, start-to-end frame interpolation (F2's, FunInp, FLF2V) - all with separate V2V option

• Control Loras for 1.3b - tile and depth

• FunControl - to restyle video based on input image and it's ControlNet

• Wrapper IV2V node - makes it possible to upscale to resolutions unavailible with native nodes; can be used as v2v after converting to nodes and removeng Start Frame related nodes

• Skyreels V2 DF - seamless extention, vram-hungry, even though it's a wrapper

• LTX infinite keyfame interpolation - neat feature that Wan lacks

• HunYuan integration - T2V and I2V with corresponding V2Vs - to use with speedy 5-step AccVideo and HY loras, best performance with T2V.

Wrapper:

• basic functions, V2V anything

• FantasyTalking - lipsync tool

• VACE (KJ's implementation) is a framework that allows to transform T2V or any finetunes into all-in-one models capable of I2V, LF2V (last frame to video), FLF2V (Interp), Control (CN maps or Dot), Reference, Video outpainting, I2V outpainting, and probably more. Native implementation is not as versatile since it is basically VACE merged with T2V ckpt. All it's features can be used in any combination, even wacky V2V with control to LastFrame, just toggle things correctly.

• Phantom (only 1.3b availible at the moment) is a tune that allows up to four references of object, characters or backgrounds, kinda same as Flux's UNO and InstantCharacter, but in one, and in video, and in 1.3b meaning lower vram reqs. Thanks to KJ's VACE implementation they can be paired - using Phantom with VACE Reference and I2V outpaint produces spectacular results (only at 1.3b!) for one and two objects, but starts struggling with 3. VACE in that case should be adjusted to start at ~.4 with reduced strength to allow Phantom do it's thing.

Changelog

v2.7 17.05.2025

WanWrapper workflow updated

• VACE 14b and 1.3b

• Phantom

• FantasyTalking

• CausVid (3step) lora

• Spline dot control example

v2.6 24.04.2025

• Wrapper IV2V node - crank up blocks to swap and upscale to XL resolution

• FLF2V - official start-to-end frame model that rocks

• Skyreels V2 DF - video extention model that takes last frames of provided video and seamlessly extends it - pretty much FramePack, but Wan

v2.5 7.04.2025

Wan wrapper consolidated - separate workflows are likely better, keeping it as proof of concept

v2.1. 4.04.2025

InterpV2V nodes added.

v2. 3.04.2025

HunYuan - standard tools: T2V, V2V, I2V (including LeapFusion), IV2V (performs poorly)

introduction of Group Nodes

v1. 31.03.2025

initial release