Late Night TV was a series about falling asleep in front of the television with your eyes open. The idea was a set of channels you drift between, half awake.

We never built the frame for it: the TV, the channel changes, the slide between waking and sleep. What exists are eight short loops, one for each channel: an action movie, a car ad, the news, a drink ad, a talk show, a late movie, a recruitment ad, and a mecha cartoon.

How they were made

Everything ran on a laptop in ComfyUI, a free tool where you build an image pipeline by connecting boxes on a canvas.

To set it up on a Mac with Apple Silicon:

git clone https://github.com/comfyanonymous/ComfyUI.git
cd ComfyUI
python3 -m venv venv && source venv/bin/activate
pip install -r requirements.txt
python main.py

ComfyUI then opens in the browser at localhost:8188. Add ComfyUI Manager to the custom_nodes/ folder. It lets you install everything else from a search box.

Animation needs three more add-ons: AnimateDiff-Evolved, VideoHelperSuite, and Advanced-ControlNet. The models come from Hugging Face. Each kind goes in its own folder: models/checkpoints, models/loras, models/animatediff_models, and models/animatediff_motion_lora.

Keep a list of which model made which clip. A text file is enough.

The settings

• Model: Stable Diffusion 1.5 with the AnyLoRA checkpoint, which gives the drawn anime look.
• Motion: AnimateDiff, with the v3_sd15_mm.ckpt motion module.
• Length: 16 frames at a time, with uniform context options and pyramid scheduling.
• Camera moves: optional motion LoRAs for pan, zoom, tilt, and rotate.
• Output: a Video Combine node saves the mp4.

All eight channels came from the same setup. Only the scene changed.

Other tools

For a quick look at an idea, DiffusionBee is easier. It is one installer and needs no Python. To fix the seed and change one setting at a time, AUTOMATIC1111’s WebUI shows every option in the browser. ComfyUI fits best when you want to save the whole process and run it again.

Why run it on a laptop

When it runs on my own machine, I know what made each clip: the model, its license, and the settings. I can save the whole setup as one file, hand it to someone else, and make the same clip again later.