A Python script that downloads all model files required by the official built-in workflow templates in ComfyUI, placing each file directly into the correct subfolder under ComfyUI/models/.
All URLs are sourced from and verified against the official Comfy-Org/workflow_templates repository.
- ✅ Downloads directly to the correct
models/subfolder — no manual moving - ✅ Skips files that already exist (safe to re-run)
- ✅ Colour-coded terminal output (green = success, blue = skipped, yellow = warning, red = failed)
- ✅ Resume-safe — uses
.parttemp files with atomic rename on completion - ✅ Retry logic — 3 attempts with delay between each
- ✅
--verifymode — HEAD-checks all URLs without downloading anything - ✅
--dry-runmode — shows what would be downloaded without doing it - ✅ Interactive menu or fully CLI-driven
pip install requests tqdm1. Set your ComfyUI root path at the top of the script:
COMFYUI_ROOT = r"C:\ComfyUI" # Windows
# COMFYUI_ROOT = "/home/user/ComfyUI" # Linux / Mac
# COMFYUI_ROOT = r"D:\AI\ComfyUI_windows_portable\ComfyUI" # Portable2. (Optional) Set your HuggingFace token for gated models:
HF_TOKEN = "hf_xxxxxxxxxxxxxxxxxxxxxxxxxxxx"Get your token at huggingface.co/settings/tokens.
3. Run it:
python download_comfyui_models.pypython download_comfyui_models.py # interactive menu
python download_comfyui_models.py --all # download everything
python download_comfyui_models.py --required # skip optional LoRAs/alternates
python download_comfyui_models.py --category wan21 # one category only
python download_comfyui_models.py --verify # HEAD-check all URLs, no download
python download_comfyui_models.py --dry-run # preview without downloading
python download_comfyui_models.py --list # print all files and exit| Category key | Templates covered |
|---|---|
get_started |
Z-Image Turbo (T2I), Qwen-Image Edit, Wan 2.2 I2V, Hunyuan3D 2.1 |
flux1 |
FLUX.1 Dev / Schnell / Fill / Kontext + ControlNets |
flux2 |
FLUX.2 Dev + FLUX.2 Klein |
z_image |
Z-Image Turbo LoRA |
qwen |
Qwen-Image T2I + Edit + InstantX ControlNet |
other_image |
Ovis-Image |
sd_classic |
Stable Diffusion 1.5 + Canny ControlNet |
sdxl |
Stable Diffusion XL + Refiner |
sd3 |
Stable Diffusion 3.5 Large + Medium |
wan21 |
Wan 2.1 T2V + I2V (14B FP8) |
wan22 |
Wan 2.2 T2V + I2V (14B FP8, MoE high+low noise pairs) |
hunyuan_video |
HunyuanVideo T2V + I2V (BF16) |
ltx |
LTX Video T2V + I2V |
cosmos |
Cosmos Predict2 |
mochi |
Mochi Preview |
3d |
Hunyuan3D 2.0 + 2.1 |
audio |
ACE-Step v1, Step-Audio |
upscale |
4x-UltraSharp, RealESRGAN x4+ |
Some models require you to accept a licence agreement on HuggingFace before your token will work. Visit each link and click Agree while logged in, then your HF_TOKEN will unlock the download.
| Model | Licence page |
|---|---|
| FLUX.1 Dev | black-forest-labs/FLUX.1-dev |
| FLUX.1 Fill Dev | black-forest-labs/FLUX.1-Fill-dev |
| FLUX.1 Kontext Dev | black-forest-labs/FLUX.1-Kontext-dev |
| Stable Diffusion XL | stabilityai/stable-diffusion-xl-base-1.0 |
| Stable Diffusion 3.5 Large | stabilityai/stable-diffusion-3.5-large |
| Stable Diffusion 3.5 Medium | stabilityai/stable-diffusion-3.5-medium |
Models showing
[401]in--verifyoutput that are not in the table above only require a valid HF token (any logged-in account) — no special licence agreement needed.
Downloading every unique model file requires approximately 250–350 GB of disk space. The get_started category alone covers the four default new-template subgraphs and is a good starting point (~60 GB).
This script covers all local / non-API built-in templates — the ones that run entirely on your own GPU. It does not cover cloud API-only templates (FLUX.1 Pro, Runway, Kling, Wan 2.6 API, etc.) as those require paid API keys and have no local model files.
- Wan 2.2 uses a Mixture of Experts (MoE) architecture. Both the
high_noiseandlow_noiseexpert model files are required — each template loads them into separate nodes simultaneously. - Mochi T5 encoder is saved as
t5xxl_fp8_e4m3fn_mochi.safetensorsto avoid filename conflicts with other T5 encoders in the same folder. - URLs are verified against live HuggingFace repositories as of March 2026. Model repos are occasionally reorganised — use
--verifyto check for broken links before a large download session.
If you find a broken URL or a missing template, please open an issue or PR. The model list is in the MODELS array near the top of the script and is straightforward to update.
MIT
