Gemma 4 and ComfyUI: setup and prompt node tutorial

This tutorial sets up Gemma 4 in ComfyUI, then demonstrates Pixaroma nodes for managing image generation prompts. The presenter first updates ComfyUI and the custom nodes, places the adapted model in models/text_encoders, and refreshes node definitions. A text generation workflow combines a reusable formula with a short prompt to produce a longer description.

For this local LLM workflow, the presenter reports around 8 GB of VRAM use on a 24 GB card and warns that insufficient memory makes generation slow. At recording time, he says GGUF models do not work with these nodes. Unchanged inputs reuse the generated prompt; changing the seed triggers another generation. Connecting Gemma to Z image turbo keeps both models loaded and increases VRAM demand.

Other examples turn an image into a prompt, transcribe audio, and describe video frames. The video workflow extracts frames without audio, and one description test covers only four frames rather than every frame.

Prompt Stack combines enabled text fields. Prompt Pack queues prompts separated by lines or blank lines, while Prompt Multi routes indexed prompts or queues individual fields. Prompt Reader retrieves prompts from generated images, though some custom workflows may escape detection. Later sections cover text overlays, background removal models, node colors, and optional button effects.