🎯 Compact Wan Workflow — Simplify Your Setup 🚀 — v1.1

Author notes from 🔨Wan Nodes.json · selected release

Upscaler for video using Wan 2.2 5B. First, upscale the input video using a regular upscaling model. Then, you can normalize the video to reduce its resolution to the required size. The node already contains a basic prompt for upscaling, but you can replace it with your own via ?positive and ?negative. If the final video quality is unsatisfactory, try reducing the denoise parameter. The higher the denoise, the more the final video differs from the original.

Author notes from 🔨Wan Nodes.json · selected release

Calculates the number of frames for Wan per second and the generated video's fps

Author notes from 🔨Wan Nodes.json · selected release

This node helps work with long videos by breaking the video into many small pieces - batches. Also, at each iteration, you can add not only images but also other types to a common buffer. The result will be a complete buffer of the type passed to "acc". You can find an example of using this node in the "Examples" section.

Author notes from 🔨Wan Nodes.json · selected release

Wan Setup allows you to install the necessary models for working with Wan. Wan Optimizer enables video generation optimization and includes some nodes to improve generation quality. Setup occurs within the node itself, so click the button to enter the subgraph to edit the setup according to your available models and hardware.

Author notes from 🔨Wan Nodes.json · selected release

This group contains nodes that can help when working with Wan Animate. These nodes greatly simplify the creation of avatar or animation generation algorithms. Here is a description of each node: - Wan Animate Resize Image & Video - changes the size of images and videos and aligns them to a single dimension. This helps use reference_image and control_video of the same size for better generation control and artifact removal in the final video. You can also set normalization to optimize video generation for your device. - Face Estimator - detects the character's face on a video. - Pose Estimator - detects the character's body and arms on a video. - Character Estimator - a basic node for isolating a character from a general video. You can use this node when you want to specify your character mask. You can reduce the block_size parameter to get a more accurate character isolation, but this may increase VRAM consumption during video generation through Wan Animate. - Points Character Estimator - a simpler node to use. Allows creating a character mask based on points on an image. Use it together with the Points Editor from kjnodes. This node inherits from Character Estimator. - Auto Character Estimator - a simpler node to use. Allows creating a character mask based on a hint. Specify who you want to isolate on the video, and the node will automatically generate a mask for you. This node inherits from Character Estimator.

Author notes from 🔨Wan Nodes.json · selected release

This group contains a huge number of samplers that can be used in various situations. The nodes are arranged such that the top group contains standard nodes without any modifiers. In the lower group, modifiers for standard samplers are located. Often these are Lightning modifiers, but there are also ones that improve quality without compromising video generation speed: - Wan 2.2 14B LightX2V Advanced KSampler - an enhancement to the Wan 2.2 14B LightX2V KSampler that includes an additional step for high models without Lightning LoRA, improving character movements and actions on video. - Wan 2.2 14B LightX2V+Pusa KSampler - improved video quality achieved through the use of Pusa LoRA. LoRA is used for some samplers, so please set your path to it.

Author notes from 🔨Wan Nodes.json · selected release

Image normalization allows you to adjust the input image to match a specific model. At its core, it uses the "Scale Image to Total Pixels" node. Try to avoid using this node for Wan 5B, as the model performs poorly with resolutions other than 1280x704 or 704x1280. Additionally, this node helps process images for low VRAM devices. The following sections will specify appropriate normalization settings for certain models, as well as recommendations for low VRAM devices. For Wan 14B: - Low VRAM: 0.21 meg - High VRAM: 0.50 meg Resolution ratio reminder: 1080p - 2,074 meg 720p - 0.9216 meg 480p - 0.4096 meg My config for 6GB VRAM: 1. For Wan 14B: - Maximum - 0.32 meg for 5s/16fps - Stable - 0.25 meg for 5s/16fps 2. For Wan 5B: - Maximum - 0.86 meg for 5s/24fps - Stable - 0.8 meg for 5s/24fps

Author notes from 🔨Wan Nodes.json · selected release

# Wan Nodes Workflow ## Description **This workflow was designed to simplify creating your own projects for Wan using pre-built, advanced, and most importantly, compact nodes.** Although this workflow was created for use with Wan’s generative model, some of the nodes can be utilized outside of Wan as well. The author aimed for the idea of "keeping the backend far from human eyes," so all logic was hidden within separate nodes that can be used in your own projects. Additionally, the author tried to simplify project development by adding auxiliary improvements to each node, such as optimizations and normalizations to ensure workflows can run on low-end devices. Regarding low-end devices, the author left recommendations for optimizing the Wan pipeline in the section "Optimization Recommendations". > Warning: > When copying nodes, parameter shifts may occur due to quirks (bugs) in ComfyUI. The nodes in this workflow have been tested and will not break when moved. To achieve this result, nested nodes were unpacked. All nested nodes were grouped to make them easier to identify. Please follow this rule to ensure stable node copying. ## Designation ### Node Parameters - ?input - optional input - _output - hidden output (often used for debugging) - [input/output] - input/output within a single iteration ### Node Colors - yellow - utility (auxiliary) nodes. - purple - nodes for initial setup and pipeline configuration - cyan - condition nodes. - green - sampler nodes. - red - encoders/decoders - black - utility (auxiliary) nodes, but for Wan Animate (made to highlight significance, but in reality these are the same auxiliary nodes, lol XD). - blue - compact "all-in-one" nodes. Responsible for specific tasks. ## Custom Nodes - [ComfyUI-GGUF](https://github.com/city96/ComfyUI-GGUF) - [ComfyUI-wanBlockswap](https://github.com/orssorbit/ComfyUI-wanBlockswap) - [ComfyUI-MagCache](https://github.com/Zehong-Ma/ComfyUI-MagCache) - [rgthree-comfy](https://github.com/rgthree/rgthree-comfy) - [ComfyUI-KJNodes](https://github.com/kijai/ComfyUI-KJNodes) - [ComfyUI-Easy-Use](https://github.com/yolain/ComfyUI-Easy-Use) - [comfyui_controlnet_aux](https://github.com/Fannovel16/comfyui_controlnet_aux) - [ComfyUI-VideoHelperSuite](https://github.com/Kosinkadink/ComfyUI-VideoHelperSuite) - [ComfyUI-Frame-Interpolation](https://github.com/Fannovel16/ComfyUI-Frame-Interpolation) - [ComfyUI-segment-anything-2](https://github.com/kijai/ComfyUI-segment-anything-2) - [ComfyUI-SAM2](https://github.com/neverbiasu/ComfyUI-SAM2) ## Model Links: 1. ComfyUI Repackaged: - [https://huggingface.co/Comfy-Org/Wan_2.2_ComfyUI_Repackaged](https://huggingface.co/Comfy-Org/Wan_2.2_ComfyUI_Repackaged) 2. Calcuis GGUF Repackaged: - [https://huggingface.co/calcuis/wan2-gguf](https://huggingface.co/calcuis/wan2-gguf) 3. Kijai Repackaged: - [https://huggingface.co/Kijai/WanVideo_comfy](https://huggingface.co/Kijai/WanVideo_comfy) ## Optimization Recommendations Below is a table of parameters that can reduce VRAM consumption or speed up video generation: | Node | Parameter | Impact | Performance | Feature | | ---- | ---- | ---- | ---- | ---- | | Wan Setup->Load Wan | GGUF | strong | enable | reduces VRAM, speeds up generation | | Wan Setup->Load Clip | GGUF | medium | enable | reduces VRAM, speeds up generation | | Wan Optimizer | Sage Attention | strong | auto | reduces VRAM, speeds up generation | | Wan Optimizer | FP16 | low | enable | reduces VRAM, speeds up generation | | Wan Optimizer | MagCache | medium | enable | speeds up generation | | Wan Optimizer | Compile | medium | enable | reduces VRAM, speeds up generation | | Wan Optimizer | Block swap | strong | higher - better | reduces VRAM | | Image Normalize (and other nodes that use this) | is_scale | strong | enable | reduces VRAM, speeds up generation | | Image Normalize (and other nodes that use this) | megapixels | strong | lower - better | reduces VRAM, speeds up generation | | Decode | VAE Tiled Decode | strong | enable | reduces VRAM | Best settings for low VRAM (6-8 GB): - Wan Setup->Load Wan - GGUF - Q4_K_M - Wan Setup->Load Clip - GGUF - Q4_K_M - Wan Optimizer - Sage Attention - auto - Wan Optimizer - FP16 - Wan Optimizer - Block swap - 40, if working with high-resolution videos or Wan Animate. - Image Normalize - megapixels - 0.21 for Wan 2.2 14B - Decode - VAE Tiled Decode Additionally, I recommend setting virtual memory in case RAM fills up. You may also use Mem Reduct if RAM is extremely limited. ## My setup: - Laptop: Asus TUF Dash F15 - GPU: Nvidia RTX 3060 6GB Laptop - CPU: Intel i5-11300H - RAM: 24 GB + 32 GB swap ## License *Why? Use at your own risk!* ¯\_(ツ)_/¯ ## About This workflow was created by Bleyn. All social media links and donations: - [CivitAi](https://civitai.com/user/NeuroContent) 🧠 - [Instagram](https://www.instagram.com/neurocontent10/) 🎨 - [Pinterest](https://pinterest.com/NeuroContent10/) 📌 - [Donate](https://dalink.to/neurocontent) 🍵

Author notes from 🔨Wan Nodes.json · selected release

Calculates the difference in steps between high and low models.

Author notes from 🔨Wan Nodes.json · selected release

Эта нода позволяет использовать CLIP Vision Encode, даже если clip_vision отсутствует. Если clip_vision не существует (bypass или просто не подключен), то clip_vision_output = null.

Author notes from 🔨Wan Nodes.json · selected release

The most powerful node among all the ones presented. This node simplifies the use of the full pipeline for Wan Animate, enabling work with long videos without constantly copying the node for each video segment. It works thanks to the While Loop in Easy-Use. The algorithm generates each video segment (batch) as many times as needed to process the entire video. After obtaining each batch, the algorithm combines them together. The last batch may extend beyond the video length (for example, in the final segment of the video, 15 frames remain, while the batch processed is 45 frames). This does not interfere with video generation, but it may result in a motion stop at the video boundary. Therefore, the algorithm cuts out the excess generation when sending the result. However, you can still observe it in the hidden output _video_batch. You can find an example of using this node in the "Examples" section.

Author notes from 🔨Wan Nodes.json · selected release

These are fully ready nodes that can be used in any workspace. It includes Wan Setup + Wan Optimizer + Loras. The setup occurs directly within the node, so press the button to enter the subgraph to edit the setup according to your models and hardware. If you want more control over the setup, I recommend using smaller nodes from the Wan Nodes->Setup section.

Author notes from 🔨Wan Nodes.json · selected release

Processing...

Author notes from 🔨Wan Nodes.json · selected release

Processing...