Option to Skip Symlink Scanning for Network Paths

[oumad]

Skip Symlink Scanning for Network Paths

Problem

When a symlink in the lora folder points to a network path containing many LoRA files, ComfyUI experiences severe performance degradation. Workflows that normally take 5-10 seconds can take 11+ minutes because ComfyUI attempts to recursively scan and list all files in the symlinked network directory before executing workflows, even when using the API with direct file paths.

Solution

Added a --skip-symlink-scan command-line flag that prevents ComfyUI from following symlinks during directory scanning operations. This allows:

• Fast startup/workflow execution: Symlinked network directories are skipped during scanning
• Direct file access still works: When providing file paths directly via a node or API json (e.g., symlink/SDXL/my_lora.safetensors), get_full_path() can still resolve and access files through symlinks because it uses direct file existence checks rather than directory scanning

Changes Made

1. Added CLI Argument (comfy/cli_args.py)

• Added --skip-symlink-scan flag to control symlink following behavior during directory scanning

2. Modified recursive_search() (folder_paths.py)

• Checks for --skip-symlink-scan flag
• When enabled, filters out symlinked directories during os.walk() traversal
• Prevents scanning into network paths accessed via symlinks

3. Modified recursive_search_models_() (app/model_manager.py)

• Applies the same symlink skipping logic for model file listing operations
• Ensures consistency across all directory scanning operations

Usage

Run ComfyUI with the flag:

python main.py --skip-symlink-scan

This prevents the performance slowdown while still allowing filen paths to work. Note that dropdown lists in the frontend will not show files from symlinked directories when this flag is enabled, but direct API usage is unaffected.

Technical Details

• get_full_path() uses os.path.isfile() which automatically resolves symlinks, so direct file access continues to work
• Directory scanning uses os.walk() with followlinks=False when the flag is set
• Symlinked directories are detected using os.path.islink() and filtered out before traversal

[guill]

Skipping the symlinks when scanning kind of seems like a roundabout way to hide the symptoms of the issue for this specific use-case rather than a true fix. This same problem can happen without symlinks when using a mounted directory like sshfs or a samba share.
I'd prefer to see something that just caches the list of available models to prevent the frequent re-scanning rather than side stepping this very specific use-case.

[oumad]

Skipping the symlinks when scanning kind of seems like a roundabout way to hide the symptoms of the issue for this specific use-case rather than a true fix. This same problem can happen without symlinks when using a mounted directory like sshfs or a samba share. I'd prefer to see something that just caches the list of available models to prevent the frequent re-scanning rather than side stepping this very specific use-case.

I agree that this solution while it provides a workaround (create symlink for paths that we don't desire to scan), it is still inherently limited.
So I'm thinking of making a different PR maybe.
ComfyUI seems to already cache model file lists using cache_model_file_list_ and cached_filename_list_, it's just it doesn't trust that cache and it scans again everytime we run a workflow. So would an argument that force it to "trust" the first scan be worthwhile ? which means a restart would be required to refresh that list, unless we provide some sort of polling interval.