WHY?! Wan and HunYuan.. get it, eh?

v2.1. 4.04.2025

Wan InterpV2V nodes added, Flow2 and FunInp variants. Easy loops

Fixed Interp nodes, it appears you gotta pass both frames into Clip Vision Encode

Small changes inside group nodes to reduce connections drag

v2. 3.04.2025

What's new:

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

• mini-workflows from v1 are now Group Nodes - 14 overall, including HY

• Sigmas split trick for Wan - run some steps on CFG=1 to speed up generation at cost of details, detached from Group Nodes - use separately

• Refiner - if you're happy with motion, but not with quality, run some additional steps (increasing steps in sampler almost guaranteed to change motion)

HY brings to the table it's speed and Loras. You can run a bunch of gens with it to cherry pick and refine with Wan.

v1 was unweildy, Group Nodes solve this issue somewhat, reduce node gore, reduce used space and are overall easier to use. But what a pain to set them up. And I already have ideas of improvements that will lead me to recreating each one -_-

Changes:

• dropped 'WanVideo Enhance A Video (native)' node completely - it was reducing background quality by alot while improvements to main object weren't noticable, and also it's highly unstable (no more people in latex hoods instead of hair, ha!)

• removed Everything Anywhere nodes, since now there's HY, or was it Anything Everywhere? I mix them up constantly

• instead of AE nodes are Pipes from Crystools

• replaced gargantuan Sampler farm group with single 'KSampler Clone' Group Node - build your own farm if you want

What's next?

• KJ's wrappers

• see if HunYanLoom worth it

• sound generation?

• whatever else new tricks come up that I fail to keep up with

v1. 31.03.2025

This is an entire collection of workflows designed to be combined in any way.

It is complex, but every tool is akin to a node with input parameters, output video and singular purpose. It is a toolkit to churn out long videos consisting of shorter clips generated with separate prompts, lengths and lora stacks, and it's rather compact for what it does.

It features:

• organized checkpoint profiles with optimizations and quality upgrades applied;

• structured node groups with singular function;

• easy setting of video resolution and length.

Workflow contains notes with explanations and divided into four loose areas.

1. Models

Every checkpoint has it's own loader to avoid constant parameter swapping. KJ's Set/Get nodes are used to access checkpoints.

Applied optimizations:

• Sage Attention and TorchCompile by KJ, require triton - pure speed boost;

• Tea Cache - speed boost at cost of quality.

Applied quality improvements:

• SLG by KJ, should be used with TeaCache;

• CFGZeroStar - native.

Clips and VAE are set to global with Anything Everywhere nodes.

2. Staging area

Here you can:

• define FPS (default 16) for automatic calculation of video length;

• enter default prompts for Anything Everywhere;

• set video base and upscale sizes in few clicks, unless you want non-standard SD size;

• load image for i2v.

3. Toolkit area

The bread and butter of this workflow. It consists of self-sufficient mini-workflows.

These are:

• T2V;

• V2V;

• I2V;

• IV2V - not style transfer (aka I+V2V), just combined I2V and V2V;

• Upscale - basic with upscale model, best to be followed with V2V or IV2V;

• Draw text;

• Tiled Enchance - improves quality of a video using Tile Control lora;

• Depth motion control - utilizes Depth Control lora, can be used to restyle video (when you want to mimic motion with character lora, for example);

• LTX Frame Interpolation - utilizes LTX model to create video with unlimited amount of keyframes;

• Wan Frame Interpolation - by Flow2, interpolates between start and end frame, uses I2V model;

• Wan FunInpaint Interpolation - native, also interpolates between start and end frame, uses WanFun models;

• Wan FunControl Style Transfer - native, actual I+V2V, uses WanFun models;

• Sampler farm - 9 samplers to seed hunt while afk.

For these groups/blocks it is required to:

• enter prompt and video length in seconds, where needed - nonV to V

• pass model and/or starting image, reference image, video, positive conditioning.

Node 'WanVideo Enhance A Video' by KJ is applied wherever possible, grants quality boost. It is set up here and not in Models area because it works with latent. It sometimes causes errors and should be bypassed in that case.

Every block is supplied with usage instructions. Same with few utility nodes above Staging area.

4. Canvas

Remaining empty space is where you can drag, copy and chain and mix every tool in any combination.

Most convenient way is to save these blocks as node templates.

Refer to examples for particular use cases.

It took quite an effort to compose this workflow, so I hope you will not be rebuffed by its initial complexity and extract real value from it.