A ComfyUI extension that scans your loaded workflow for missing model files, finds them on HuggingFace and CivitAI, and downloads them straight into the correct folder. Saves disk space by hardlinking copies you already have.

Deutsche Version: README.de.md

⚠️ Disclaimer

This project was vibe-coded. It was built iteratively in a conversation with an AI coding assistant - shipped because it works for the author, not because every code path was reviewed line by line.

Use at your own risk. The author provides this software AS IS, without warranty of any kind. The author accepts no liability for anything that happens as a result of running this code, including but not limited to: deleted or corrupted files, downloads to the wrong location, exhausted disk space, leaked API tokens, broken ComfyUI installations, unexpected bandwidth or storage costs, or any other direct or indirect damage.

The plugin moves files, deletes the .part of incomplete downloads, creates hardlinks/symlinks across your models/ tree, and stores API tokens in a local config.json. Back up anything you cannot afford to lose before you run it for the first time. Try it on a test ComfyUI install first if you have one.

See the LICENSE (MIT) for the formal terms.

• One-click scan of the loaded workflow detects every referenced model file - across all custom-node packs - and tells you which ones are missing.
• Smart folder routing uses ComfyUI's own folder_paths, your extra_model_paths.yaml, custom-node-registered folders (e.g. Kijai's detection/ for ViTPose / YOLO), and filename heuristics to pick the right destination. Subfolder paths from the workflow (Wan2_2/lightx2v/...) are preserved.
• Multi-strategy search:
  1. Curated database of well-known models (FLUX, SDXL, ControlNet v1.1, ...)
  2. HuggingFace repo-name search
  3. HuggingFace full-text README search (catches files referenced in READMEs even when the repo name doesn't match the filename)
  4. Fallback list of well-known umbrella repos (Kijai/WanVideo_comfy, Comfy-Org/frame_interpolation, lightx2v/Wan2.2-Distill-Loras, ...)
  5. CivitAI search
• Background downloader with progress bar, smooth animation, download speed and ETA, resume support, cancel, and HTTP-error detection (catches the "server returned an HTML login page instead of the model" trap).
• Disk-space saver: when a model with the same name + size already lives anywhere in your models/ tree, the plugin can hardlink (or symlink) it instead of downloading again. Optional, opt-in.
• Move existing: shifts files that landed in the wrong folder into the right one, no re-download needed.
• Duplicate protection: clicking download a second time on a model that's already in flight produces a clear toast notification instead of starting a parallel transfer.
• API tokens for HuggingFace gated repos (FLUX.1-dev) and CivitAI models. Tokens are stored in a local config.json and never sent back to the browser - only a masked preview (hf_xx••••wxyz) is shown.

Clone (or copy) the repository into your ComfyUI custom nodes folder:

cd ComfyUI/custom_nodes
git clone https://github.com/<your-username>/comfyui-modelsearchandload.git

Restart ComfyUI. A new Models tab appears in the sidebar (download icon). On older ComfyUI builds without the sidebar API, a floating "Models" button shows up in the bottom-right corner.

No additional Python packages are required - the plugin uses only the standard library (urllib, threading, json, ...) and ComfyUI's existing aiohttp server.

Tested on Windows and Linux. All paths use os.path.join, all string comparisons use os.path.normcase, and link operations fall back gracefully when the OS doesn't support them (e.g. Windows symlinks without Developer Mode → falls back to hardlink → falls back to copy).

1. Load a workflow in ComfyUI.
2. Open the Models sidebar tab.
3. Click Scan workflow - missing models appear in a list with their target folder path.
4. Click Find sources on a single entry, review the candidates, and hit Download on the source you want to use.
5. Watch progress in the Downloads panel below. Each row shows the live status, percentage, speed, and ETA. Finished files animate away from the missing list.

Other useful actions:

• Move existing scans your models/ tree for files that match a missing entry by name, then moves them to the location ComfyUI expects (correct folder + subfolder).
• Settings opens a modal for HuggingFace / CivitAI API tokens and the disk-space (linking) options.

When the same model is referenced from two different paths (e.g. Wan2_2/foo.safetensors and foo.safetensors), or when the plugin needs to satisfy a download for a file that's already on disk elsewhere, it can create a filesystem link instead of downloading the data again.

Open Settings → Disk space, tick Reuse existing files via filesystem links, and pick a mode:

The plugin matches by filename + file size. SHA256 verification is not done because reading multi-GB models would be too slow; the size check is good enough in practice for distinct model versions.

When linking succeeds, the job status reads ✓ Hardlinked or ✓ Symlinked and no bytes are downloaded.

Some models require authentication:

• HuggingFace token for gated repos like black-forest-labs/FLUX.1-dev. Get one at https://huggingface.co/settings/tokens (a read-only token is enough). You also need to accept the model's license on the HuggingFace web UI before the download will work.
• CivitAI API key for many CivitAI downloads. Get one at https://civitai.com/user/account.

Open Settings, paste the token, click Save. After saving, the input clears and the status line shows ✓ Stored: hf_xx••••wxyz. The real value never leaves the server. Click Clear to remove a stored token.

The plugin registers these endpoints on the running ComfyUI server, so external scripts can drive it too:

The download endpoint returns {"job": {...}, "duplicate": true} when a job for the same filename + destination is already running, so clients can avoid spawning parallel transfers.

Edit known_models.json to teach the plugin about a custom model:

{
  "my_favourite_lora.safetensors": {
    "folder": "loras",
    "url": "https://huggingface.co/user/repo/resolve/main/file.safetensors",
    "source": "huggingface",
    "size": 134217728
  }
}

Folder keys follow the ComfyUI convention: checkpoints, loras, vae, controlnet, clip (= text_encoders), clip_vision, diffusion_models, upscale_models, embeddings, style_models, ipadapter, gligen, hypernetworks, frame_interpolation, detection, ...

Pull requests with additions to the curated database are welcome.

• HuggingFace search blind spots: very obscure filenames that appear neither in any repo name nor in any README cannot be found automatically. For these, paste the URL directly into the candidate's search modal or add the model to known_models.json.
• Gated models require both a token and an accepted license on the HuggingFace web UI.
• CivitAI rate limits - bulk-downloading lots of CivitAI models in quick succession can hit their API rate limit.

PRs welcome. Please:

• Run python -c "import ast, pathlib; [ast.parse(p.read_text(encoding='utf-8')) for p in pathlib.Path('.').glob('*.py')]" before pushing to make sure all Python files parse.
• Validate the JS file with node --check web/model_downloader.js (the file is written as a plain ES module).
• Don't commit config.json (it's in .gitignore) - that file holds per-installation tokens.

MIT