FunArt offers deep integration with VisionPlaid, including out-of-the-box sample workflows and dependencies. This topic describes how to create a ComfyUI project, select the VisionPlaid acceleration engine, run a sample workflow, and explains the parameters for VisionPlaid ComfyUI nodes.
Overview
FunArt provides deep integration with VisionPlaid, including the VisionPlaid acceleration engine and prebuilt workflows, models, datasets, and dependencies. You can use VisionPlaid out of the box. VisionPlaid delivers state-of-the-art end-to-end inference performance for models such as Qwen-Image-Edit-*, Qwen-Image-*, and Wan2.2. This topic covers the full process—from creating a project and running a sample workflow to explaining node parameters. For performance data and comparison charts, see VisionPlaid release notes.
Prerequisites
Before you begin, ensure the following conditions are met:
You have logged on to the FunArt console.
RAM authorization: You have activated Function Compute and related services, and granted permissions to your account. For detailed steps, see Authorize a RAM user to use image generation projects.
Account balance: Maintain an account balance of at least CNY 100 for a better experience. New users can use a free trial package to save costs.
Create and configure a ComfyUI project that uses VisionPlaid
Create a ComfyUI project in FunArt and select VisionPlaid for the Acceleration Engine in the third step of the creation process. This enables VisionPlaid acceleration. The steps are as follows.
Step 1: Select a project type
In the navigation bar of the FunArt console, click Projects, then click Create New Project. In the first step, Select Project Type, choose ComfyUI.
Step 2: Select a GPU card type and specifications
In the second step, Select GPU Card Type and Specifications, choose a GPU card type as needed. When using VisionPlaid, select an Ada-* system or a Blackwell series system. Configure specifications such as GPU memory, vCPU, memory, and disk space. Then, click Next.
Step 3: Configure other project properties and select VisionPlaid
In the third step, Configure Other Project Properties, configure the following:
Project Name: Required. Enter a project name.
Region: Select the region where the project will be deployed, such as China (Hangzhou).
Acceleration Engine: From the Acceleration Engine drop-down list, select VisionPlaid. This option configures the acceleration tool for ComfyUI. Selecting VisionPlaid enables inference acceleration for the project.
Storage Configuration: Select an option as needed. For example, Auto Configure uses Apsara File Storage NAS (NAS) to store resources such as models and plugins.
After you complete the configuration, click Next.
Step 4: Confirm and create
In the fourth step, Confirm and Create, review the configuration information—such as project name, project type, region, and GPU specifications—along with the billing details for deployed resources, such as Function Compute (FC), Apsara File Storage NAS (NAS), and Simple Log Service (SLS). After you confirm the information, click Confirm Deployment. The system begins deployment. A message appears indicating that the project is being prepared and will take three to five minutes. No GPU fees are incurred during deployment. After deployment completes, you can enter the project development stage to use the workstation for image generation and debugging.
GPU usage is billed on a pay-as-you-go basis. The estimated cost shown on the interface is for reference only. The actual cost depends on your usage. Shut down the workspace promptly after use to stop billing.
Project development and usage
After the project is created and deployed, go to the product page to develop and generate images. For information about the workstation and image generation flow, see Quick Start: Create a ComfyUI project.
Run a sample workflow
On the project page, choose Project Development > Workstation > Workflows. You can see the sample VisionPlaid workflows built into FunArt.
Select the workflow you want to run and click the Run button in the upper-right corner to start inference.
Introduction to VisionPlaid ComfyUI nodes
The following sections describe the main VisionPlaid nodes and their parameters in a ComfyUI workflow.
VisionPlaid Load Diffusion Model
unet_name: The model weights to load. This parameter functions the same as Load Diffusion Model in ComfyUI. If you use W4A4 quantization, specify the corresponding weights here. If you enable four-step optimization, download the dedicated weights for four-step optimization.
weight_dtype (Quantization Type): default behaves the same as Load Diffusion Model in ComfyUI. int4 specifies W4A4 int4 quantization. nvfp4 specifies W4A4 nvfp4 quantization. Other values behave the same as Load Diffusion Model in ComfyUI.
Diffusion Boost V1
Diffusion Boost V1 is the central entry point for controlling and configuring inference acceleration.
enable_compile: Enables or disables compilation. Using torch.compile to compile the transformer optimizes operators and improves performance.
attn_type (Attention Type): Currently supports
default,sage, andflash. `default` uses the attention implementation from ComfyUI. `flash` requires you to install FlashAttention separately. `sage` uses SageAttention, which is built into the bitforge version and requires no separate installation.enable_fbcache, residual_diff_threshold, and cache_start are not supported in the current version.
enable_offload: Enables automatic offloading of transformer weights. This reduces high memory usage during VAE decode for large images and avoids performance loss caused by unloading the transformer.
num_blocks_on_gpu: When automatic offloading is enabled, this controls how many transformer blocks remain resident on the GPU. Decreasing this value reduces the probability of memory contention with the VAE. Increasing this value may improve performance. Adjust it based on the image size.
The first run may require compilation and take longer. Performance reaches its optimum after several runs.
Parallel KSampler
In addition to standard ComfyUI parameters—such as model, seed, steps, cfg, sampler_name, scheduler, positive, negative, latent_image, and denoise—VisionPlaid provides:
num_gpus: The number of GPUs used for parallel processing. When `num_gpus=1`, inference runs on a single GPU. When `num_gpus>1`, multi-GPU sequence parallelism and communication overlap reduce communication overhead.
Enter the project: In the project list, click the ComfyUI project that you created and configured with the VisionPlaid acceleration engine to open the project details.
Use the workstation: On the Project Development tab, use the Workstation to configure models, workflows, and parameters. Set prompts and execute the workflow to generate images.
Manage files: On the File Management tab, you can view input and output files, models, and other resources.
Publish as an online service (Optional)
To publish a ComfyUI project configured with VisionPlaid as an online API for external calls, use the Online Service tab to publish and manage the configuration. For the detailed procedure, see Publish a ComfyUI project as an online service.
Do not directly modify the function configurations created by FunArt in the Function Compute console. Doing so may cause your changes to be overwritten during the next deployment.