ComfyUI Loader Utils - Optimized Model Loading
The Problem: Native ComfyUI Loading and VRAM OOM Issues
ComfyUI's native loading mechanism has a significant limitation that affects users with limited VRAM: all models are loaded into VRAM before the workflow begins execution. This approach causes several issues:
- VRAM Overflow: All models are loaded simultaneously, consuming excessive GPU memory
- Unnecessary Loading: Even models not needed for the current execution are loaded
- OOM Errors: Users with limited VRAM experience crashes and out-of-memory errors
- Inefficient Resource Management: No control over the loading order or timing
The Solution: Loader Nodes with "Any" Parameter
This custom loader module addresses these issues by:
- Flexible Node Connections: Added an optional "any" parameter to all loader nodes, allowing them to connect to any output type
- Controlled Loading Order: Users can strategically place loader nodes after other nodes, optimizing the model loading sequence
- Memory Management: Enables better VRAM management by controlling when and which models are loaded
- Sequential Loading: Models are loaded only when needed, in a controlled sequence
Features
- All standard ComfyUI loader nodes included with "_Any" suffix
- Optional "any" parameter for flexible connections
- Maintains all original functionality and parameters
- Compatible with existing ComfyUI workflows
Available Loader Nodes
CheckpointLoader_Any- Advanced checkpoint loadingCheckpointLoaderSimple_Any- Simple checkpoint loadingDiffusersLoader_Any- Diffusers model loadingunCLIPCheckpointLoader_Any- unCLIP checkpoint loadingLoraLoader_Any- LoRA model loadingLoraLoaderModelOnly_Any- LoRA model only loadingVAELoader_Any- VAE model loadingControlNetLoader_Any- ControlNet model loadingDiffControlNetLoader_Any- Diffusion ControlNet loadingUNETLoader_Any- UNET model loadingCLIPLoader_Any- CLIP model loadingDualCLIPLoader_Any- Dual CLIP model loadingCLIPVisionLoader_Any- CLIP vision model loadingStyleModelLoader_Any- Style model loadingGLIGENLoader_Any- GLIGEN model loading
Benefits for Low VRAM Users
- Reduced Memory Footprint: Load models only when needed
- Flexible Sequencing: Arrange loading order based on available memory
- Avoid OOM Errors: Control when memory-intensive models are loaded
- Improved Workflow Stability: More predictable memory usage
Usage
The "_Any" suffix nodes can be used exactly like their original counterparts, with the added benefit that they can accept connections from any node type via the optional "any" parameter. This enables better workflow design for memory-constrained environments.
Simply use these nodes in place of the standard loader nodes, and strategically connect them to control when models are loaded into memory.
Contact
- Twitter: @Lrzjason
- Email: lrzjason@gmail.com
- QQ Group: 866612947
- Wechatid: fkdeai
- Civitai: xiaozhijason
Sponsors me for more open source projects:
|
Buy me a coffee: 
|
WeChat: 
|