Generating AI Videos Locally on Apple Silicon with ComfyUI + WAN2.1

How I set up a fully local text-to-video pipeline on my Mac — no cloud GPU, no API keys, just my own machine doing the work.

Why Local Video Generation?

Text-to-video AI models have exploded in quality over the last year, but most demos point you toward cloud GPUs or paid APIs. I wanted to see how far I could get running everything locally on Apple Silicon — and it turns out, with the right model and the right setup, the answer is: surprisingly far.

For this project I used ComfyUI, the node-based interface for diffusion models, paired with WAN2.1, an open-source text-to-video model. Specifically, the 1.3B parameter variant in fp16 — small enough to run comfortably on a Mac’s unified memory, while still producing solid short video clips from a text prompt.

What You Need

• A Mac with Apple Silicon (M-series chip)
• Python 3.14
• ~25 GB free disk space for models
• ComfyUI

Step 1 — Install ComfyUI

git clone https://github.com/comfyanonymous/ComfyUI.git
cd ComfyUI
python3 -m venv .venv
.venv/bin/pip install torch torchvision torchaudio
.venv/bin/pip install -r requirements.txt

Step 2 — Download the Models

Three files, all in fp16 — keeping everything in a single, consistent precision makes the pipeline run smoothly end-to-end on Apple’s GPU backend:

# Diffusion model (~2.6 GB)
curl -L -o models/diffusion_models/wan2.1_t2v_1.3B_fp16.safetensors \
  "https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/resolve/main/split_files/diffusion_models/wan2.1_t2v_1.3B_fp16.safetensors"

# Text encoder (~11 GB)
curl -L -o models/text_encoders/umt5_xxl_fp16.safetensors \
  "https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/resolve/main/split_files/text_encoders/umt5_xxl_fp16.safetensors"

# VAE (~250 MB)
curl -L -o models/vae/wan_2.1_vae.safetensors \
  "https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/resolve/main/split_files/vae/wan_2.1_vae.safetensors"

Step 3 — Start the Server

.venv/bin/python main.py
# → open http://127.0.0.1:8188

Step 4 — Build the Workflow

The graph is straightforward — eight nodes, no custom extensions required:

UNETLoader (the WAN2.1 1.3B model) → ModelSamplingSD3 (shift = 5.0, WAN’s flow-matching schedule) → two CLIPTextEncode nodes (positive and negative prompts) → EmptyHunyuanLatentVideo (512×512, 25 frames) → KSampler (20 steps, cfg 6.0, euler / simple) → VAEDecode → SaveAnimatedWEBP

For the prompt, I went with something simple and visual:

“a fox running through a snowy forest, mountains in the background, camera tracking shot, cinematic lighting”

with a short negative prompt to steer away from blur and artifacts:

“blurry, low quality, static, watermark, text, distorted”

Step 5 — Queue It and Wait

Hit Queue Prompt. On a 25-frame clip at 512×512 with 20 sampling steps, generation finishes in a few minutes — entirely on-device.

The Result

Out came a short animated clip of a fox loping across a snowy field, mountains rising in the background — generated start to finish on my own machine, no cloud round-trip required.

you video should be at comfyUI/output/video/yourvideo.webp

What’s Next

From here, the natural next steps are playing with longer clips (more frames), higher resolutions, and different prompts to see how the model handles motion, camera movement, and scene complexity. The same workflow scales up — just adjust the EmptyHunyuanLatentVideo width/height/length and watch your memory usage as you go.

If you’re on an Apple Silicon Mac and have been wanting to try local AI video generation, this is a great starting point — no subscriptions, no cloud bills, just your own GPU cores doing the work.

Tags: #AppleSilicon #ComfyUI #AI #MachineLearning #GenerativeAI #LocalAI #MacOS #TextToVideo