How lbh-123-ai/comfyui-wan-latent-resizer works
This repository is a ComfyUI custom-node plugin for learned Wan latent resizing. The principal workflow takes an image or video latent through a version-specific node, loads a matching learned model, and returns a resized latent for downstream decoding or refinement. The README documents both Wan versions, up/downscaling, arbitrary scale factors, and image/video latent support. Source excerpts show node registration, model loading, and resizing internals; Wan2.2 details are partial, so uncertain internal wiring is omitted.
Main entry points: wan2_1_latent_Resizer_nodes.py Wan2.1 node · wan2_2_latent_Resizer_nodes.py Wan2.2 node · nodes/__init__.py Node registry · __init__.py Plugin exports
Components 12
ComfyUI Plugin
- Plugin exports
__init__.py - Node registry
nodes/__init__.py
Wan2.1 Resizing
- Wan2.1 node
nodes/wan2_1_latent_Resizer_nodes.py - Wan2.1 resizer network
nodes/wan2_1_latent_Resizer_nodes.py
Wan2.2 Resizing
- Wan2.2 node
nodes/wan2_2_latent_Resizer_nodes.py - Wan2.2 resizer network
nodes/wan2_2_latent_Resizer_nodes.py
Learned Models
- Wan2.1 weights
models/wan2.1_latent_resizer_fp16.safetensors - Wan2.2 weights
models/wan2.2_latent_resizer_fp16.safetensors
Other components
- ComfyUI workflow
- Wan image/video latent
- Resized latent
- Decode or refine
Connections 15
- ComfyUI workflow → Plugin exports: loads plugin
- Plugin exports → Node registry: exports mappings
- Node registry → Wan2.1 node: registers
- Node registry → Wan2.2 node: registers
- ComfyUI workflow → Wan2.1 node: dispatches latent
- ComfyUI workflow → Wan2.2 node: dispatches latent
- Wan image/video latent → Wan2.1 node: input latent
- Wan image/video latent → Wan2.2 node: input latent
- Wan2.1 node → Wan2.1 weights: loads weights
- Wan2.2 node → Wan2.2 weights: loads weights
- Wan2.1 node → Wan2.1 resizer network: invokes resizing
- Wan2.2 node → Wan2.2 resizer network: invokes resizing
- Wan2.1 node → Resized latent: returns latent
- Wan2.2 node → Resized latent: returns latent
- Resized latent → Decode or refine: feeds workflow