ComfyUI tutorial: Flux 2 images and LTXV video workflows

This beginner tutorial builds a ComfyUI text-to-image workflow, then adapts it for text-to-video generation. The presenter explains node inputs and outputs, prompt encoding, sampling and VAE decoding while assembling the workflow on the canvas. Installation is outside the lesson's scope; the demonstration uses Windows and ComfyUI Desktop.

A supplied download workflow retrieves the example models and their encoders. The presenter notes that the selected Flux model requires a Hugging Face login and acceptance of its license conditions. The image example uses Flux 2 with a Qwen 3 text encoder and a matching VAE. It also demonstrates a resolution selector, seed changes and an increase from eight sampling steps to 20, which the presenter judges to produce a better image.

The video example swaps in a two-billion-parameter LTXV model, T5XXL text encoder and separate VAE. An empty latent video replaces the latent image, and additional nodes assemble and save the decoded frames. The example uses 73 frames with output set to 24 fps. The presenter deliberately chooses separate model components to retain the same workflow structure.

For local AI use, the hardware caveat matters: the presenter's laptop has a 3070 Ti mobile GPU with 8 GB of VRAM. He attributes much of the slower Flux generation to the model exceeding that memory and using system RAM. These timings describe his setup, rather than a general benchmark.