Skip to content
Help ComfyUI Wiki remove ads Become a Patron
Menu and InterfaceServer Config

ComfyUI Server Configuration Documentation

Current Version: ComfyUI Desktop v0.4.5

Server configuration is done on the server side (host side) of ComfyUI Desktop, where you can set up ComfyUI’s LAN access settings, various precision settings, and cache settings, etc. When accessing from other devices on the LAN, these settings cannot be modified.

Network

  • Listen Address (listen)

    • Function: Set the IP address for the server to listen on. Usually keep default.
    • Default Value: 127.0.0.1 (localhost only access)

  • Port (port)

    • Function: Set the port number for the server to listen on. Ensure it’s not used by other applications.
    • Default Value: 8000 (like a house number, only needs changing if there’s software conflict)

  • TLS Key File (tls-keyfile) and TLS Certificate File (tls-certfile)

    • Function: Used for setting up HTTPS secure connection. Can be ignored if HTTPS is not needed.

  • Enable CORS Header (enable-cors-header)

    • Function: Allow other websites to access your server. Use ”*” to allow all websites.

  • Maximum Upload Size (max-upload-size)

    • Function: Limit the maximum size of file uploads in MB.
    • Default Value: 100 (like a suitcase capacity)

If you want to set up LAN access to ComfyUI, you need to set the listen address to 0.0.0.0, port to 8000 or another port, then other devices on the LAN can access through the host’s LAN IP address and port. For example, if the LAN IP is 192.168.1.100 and port is 8000, other devices can access ComfyUI through http://192.168.1.100:8000

This is similar to setting -listen 0.0.0.0 in the .bat file of early Portable version ComfyUI, but now it’s more convenient to set in ComfyUI Desktop.

CUDA Settings

  • CUDA Device (cuda-device)

    • Function: Choose which GPU to use
    • Simple Explanation: If you have multiple GPUs, you can choose which one to use. Like choosing which studio to work in
    • Options:
      • 0: First GPU
      • 1: Second GPU (if available)
      • null: Auto select
    • Recommendation: Users with single GPU should keep default

  • Use CUDA Malloc (cuda-malloc)

    • Function: Determine how GPU memory is allocated
    • Simple Explanation: Like deciding how to arrange workspace in the studio
    • Options:
      OptionDescriptionUse Case
      AutoLet system decide best methodRecommended for beginners
      EnableUse more aggressive memory allocationWhen need more VRAM
      DisableUse conservative memory allocationWhen having VRAM issues

Inference

Precision: Like the level of detail in painting. High precision is like painting with a fine brush - more details but slower; low precision is like using a broad brush - faster but might lose details.

  • Global Floating Point Precision (global-precision)

    • Function: Control overall computation precision
    • Options:
      OptionDescriptionUse Case
      AUTOAuto select best precision, recommended for beginnersSuitable for most cases
      FP32Highest precision mode, like using a microscopeWhen you need best image quality
      FP16Lower precision but faster, like viewing with naked eyeWhen you want faster generation

  • UNET Precision (unet-precision)

    • Function: Control precision of AI’s core painting process
    • Note: UNET is AI’s “brush”, determining how text becomes image
    • Options:
      OptionDescriptionUse Case
      AUTOAuto select, recommended for beginnersSuitable for most cases
      FP32Highest quality, but slowerWhen need best quality
      FP16Balanced mode, medium speed and qualityDaily use
      BF16Special balanced mode, for certain new GPUsSpecific GPU use

  • VAE Precision (vae-precision)

    • Function: VAE (Variational Autoencoder) handles final image detail processing, like an artist’s “coloring and refinement” technique
    • Simple Explanation: Determines the fineness of final image details, like how detailed the artist’s final touches are
    • Options:
      OptionDescriptionUse Case
      AUTOAuto select suitable precisionRecommended for beginners
      FP16Faster but might have fewer detailsWhen need quick generation
      FP32Best quality but slowerWhen need best quality
      BF16Between FP16 and FP32For specific new GPUs

  • Run VAE on CPU (cpu-vae)

    • Function: Let CPU handle final image refinement
    • Simple Explanation: Like letting the butler (CPU) do the artist’s (GPU) finishing work
    • Use Cases:
      • When GPU memory is insufficient
      • When processing very large images
      • When GPU performance is insufficient

  • Text Encoder Precision (text-encoder-precision)

    • Function: Control how precisely AI understands text descriptions
    • Simple Explanation: Like how carefully the artist understands your requirements
    • Options:
      OptionDescriptionUse Case
      AUTOAuto select best precisionRecommended for beginners
      FP32Most careful understanding, but slowerWhen need precise prompt interpretation
      FP16Quick understanding, might have slight deviationWhen need faster speed
      BF16Balance between bothFor specific GPU optimization

Memory

  • Force Channels-Last Memory Format (force-channels-last)

    • Function: Change how image data is arranged in memory
    • Simple Explanation: Like rearranging tools in the studio for better efficiency
    • Recommendation: Keep default unless you understand its impact

  • DirectML Device Index (directml)

    • Function: Select DirectML device
    • Simple Explanation: A special painting mode, mainly for AMD GPUs
    • Use Case: AMD GPU users might need to set this

  • Disable IPEX Optimization (disable-ipex-optimize)

    • Function: Disable IPEX optimization. Usually doesn’t need changing.
    • Default Value: false

Preview Settings

  • Preview Method (preview-method)

    • Function: Control how generation process is previewed
    • Simple Explanation: Like whether to see the artist’s creation process
    • Options:
      OptionDescriptionUse Case
      NoPreviewsNo previewWhen need fastest speed
      LatentShow blurry creation processWhen want to see progress
      TaesdShow clearer creation processWhen want to see clear progress

  • Preview Image Size (preview-size)

    • Function: Set preview window size
    • Simple Explanation: Like deciding sketch size
    • Recommended Values:
      • Normal use: 512
      • Low performance: 256
      • High performance: Can set larger

Cache

  • Use Classic Cache System (cache-classic)

    • Function: Use traditional cache management method
    • Simple Explanation: Like artist organizing tools and materials in fixed way
    • Use Cases:
      • When system memory is sufficient
      • When need stable performance
    • Recommendation: Keep default unless having memory issues

  • Use LRU Cache (cache-lru)

    • Function: Set amount of recently used data to cache
    • Simple Explanation: Like limiting artist to keep only recently used tools and paints
    • Use Cases:
      • When memory is limited
      • When need to balance performance and memory use
    • Recommended Values:
      • 8GB RAM: Set to 2-3
      • 16GB RAM: Set to 4-6
      • 32GB+ RAM: Set to 8-12

Attention Settings

  • Cross Attention Method (cross-attention-method)

    • Function: Control how AI understands text and converts to images
    • Simple Explanation: Like how artist understands your requirements and conceptualizes
    • Options:
      OptionDescriptionUse Case
      autoAuto select best methodRecommended for beginners
      splitSave VRAM but might be slowerWhen VRAM is insufficient
      quadTraditional way, stable but slowerWhen need stable effects
      pytorchUse PyTorch native methodWhen having compatibility issues

  • Force Attention Upcast (force-attention-upcast)

    • Function: Force using higher precision for attention mechanism
    • Simple Explanation: Like making artist think more carefully about details
    • Use Case: When generated image details are not ideal

  • Prevent Attention Upcast (prevent-attention-upcast)

    • Function: Prevent using higher precision for attention mechanism
    • Simple Explanation: Let artist work quickly without excessive detail focus
    • Use Case: When need faster generation speed

VRAM Management

  • VRAM Management Mode (vram-management)

    • Function: Control how GPU memory is used
    • Simple Explanation: Like managing studio space usage
    • Options:
      OptionDescriptionUse Case
      AutoAuto manage, like having butler organize studioRecommended for beginners
      FullUse all available spaceWhen VRAM is sufficient
      LowSave space usageWhen VRAM is insufficient

  • Reserve VRAM (reserve-vram)

    • Function: Reserve GPU memory for other programs
    • Simple Explanation: Like reserving studio space for other uses
    • Use Cases:
      • When running other GPU-intensive programs (like games)
      • Recommended value: 2-4GB depending on other program needs

General

  • Disable xFormers Optimization (disable-xformers)

    • Function: Disable xFormers acceleration optimization
    • Simple Explanation: xFormers is like artist’s quick painting technique, disabling makes artist use traditional method
    • Recommendation: Keep enabled (unchecked) unless having issues

  • Default Hashing Function (default-hashing-function)

    • Function: Choose method to check model file integrity
    • Simple Explanation: Like checking if tools are in good condition
    • Options:
      OptionDescription
      sha256Most common and secure checking method

  • Use PyTorch’s Slower Deterministic Algorithms When Possible

    • Function: Use more stable but slower random algorithms in some cases
    • Simple Explanation: Like letting artist use more traditional but stable way for random creation
    • Recommendation: Usually no need to enable

  • Enable Some Untested Optimizations That May Lower Quality

    • Function: Use experimental optimization methods
    • Simple Explanation: Try new quick painting techniques, but might affect artwork quality
    • Recommendation: Not recommended when pursuing stable effects

  • Don’t Print Server Output to Console

    • Function: Don’t show backend running logs
    • Simple Explanation: Let artist work quietly without reporting every detail
    • Recommendation: Keep disabled (unchecked) when need to troubleshoot

  • Disable Saving Prompt Data in Files

    • Function: Don’t save prompt information in generated image files
    • Simple Explanation: Don’t record creation recipe behind artwork
    • Use Case: When want to keep prompts private

  • Disable Loading All Custom Nodes

    • Function: Prevent loading all custom function modules
    • Simple Explanation: Use only basic tools, not extra special tools
    • Use Case: When want most basic stable experience

  • Log Level (log-level)

    • Function: Set system log detail level
    • Simple Explanation: Like setting how detailed artist reports work
    • Options:
      OptionDescriptionUse Case
      DEBUGShow all details including debug infoFor developers or troubleshooting
      INFOShow general info including operations and statusRecommended for daily use
      WARNINGOnly show warnings and errorsWhen only want problem alerts
      ERROROnly show errorsWhen only care about serious issues
      CRITICALOnly show fatal errorsWhen only want most serious issues

Directory Settings

  • Input Directory (input-directory)

    • Function: Set directory for input files
    • Default Value: Empty string

  • Output Directory (output-directory)

    • Function: Set directory for output files
    • Default Value: Empty string
7. File StructureTutorial Series