Image generation acceleration-Intelligent Education Platform(IEP)-阿里云帮助中心

Use DeepytorchLite nodes to accelerate image generation in ComfyUI workflows.

Overview of DeepytorchLite acceleration nodes

Deepytorch overview

DeepytorchLite nodes use Deepytorch, an acceleration library developed by Alibaba Cloud. They can increase inference speed by over 20% in most scenarios. The actual speed increase depends on your underlying compute resources, model type, and specific workflow configuration. The acceleration nodes have the following features:

Support for mainstream model structures: DeepytorchLite nodes support a wide range of mainstream model structures, such as SD1.5, SD1.5-inpainting, SD2, SDXL, SDXL-turbo, SDXL-inpainting, SDXL-refiner, SVD, and SDXL-lighting-nstep.
Dynamic size support: Deepytorch is optimized for text-to-image scenarios with dynamic input sizes.
Support for ControlNet model acceleration.
High compatibility: DeepytorchLite is compatible with mainstream plug-ins.

Industry comparison

Acceleration framework	Kernel optimization capability	Dynamic size support	Compile-free capability	Engine file	Plug-in compatibility
Deepytorch	10% to 20%+ faster than Xformers	Support	Support	Very small	High compatibility
TensorRT	Industry-leading optimization	Supported	Not supported. It cannot be optimized in time and requires long compile times.	Large engine files. A separate engine file is generated for each model.	Many plug-in restrictions
Xformers	Excellent level. A balance between performance and compatibility.	Supported	Supported	None	High compatibility

Overview of acceleration nodes

Node types:

(1) UNet acceleration and optimization: Optimizes the model using DeepytorchLite

Node type: DeepytorchLiteOptimize

Input/Output: model

(2) VAE acceleration and optimization: Optimizes the VAE using DeepytorchLite

Node type: DeepytorchLiteOptimizeVAE

Input/Output: vae

(3) ControlNet acceleration and optimization: Optimizes the ControlNet using DeepytorchLite

Node type: DeepytorchLiteOptimizeControlNet

Input/Output: controlnet

Usage

After the environment is ready, a runnable workflow is available as shown in the following figure.

Double-click a blank area of the interface and search for Deepy. Three acceleration nodes appear in the search results. Insert the nodes into the workflow as needed.

The following shows the results of adding DeepytorchLite acceleration nodes.