Player not loading? Watch on YouTube
This August 2026 beginner ComfyUI tutorial builds a text-to-image graph from a blank canvas using Stable Diffusion 1.5. The presenter connects a checkpoint to positive and negative CLIP text encoders, a KSampler, an empty latent image and VAE Decode, then adds an image preview. The example uses a 512 × 512 canvas, 20 sampling steps and CFG 8. These are settings for the demonstrated model, rather than recommendations for every model.
The image-to-image section replaces the empty latent input with a butterfly image passed through VAE Encode. With the seed fixed, the presenter lowers denoise from 0.9 toward zero to compare how much of the source composition survives. Changing only the seed produces further variations. Preview Image supports experimentation, while Save Image adds results to the asset library; the presenter also demonstrates reopening a saved image as its workflow.
The demonstration uses Comfy Cloud. For readers interested in local AI, the troubleshooting section also discusses desktop VRAM limits and suggests reducing image resolution or choosing a less demanding model. Other examples cover incompatible model families, image-versus-latent socket types, missing inputs, unavailable file paths and outdated custom nodes. The presenter demonstrates a missing VAE connection and explains how to inspect the resulting error before reconnecting it.
