Qwen Image 2.1: ComfyUI setup and image editing tests

This tutorial covers Qwen Image 2.1 in ComfyUI for text-to-image generation and image editing. The presenter describes a 7 billion parameter model with native 2K output, RGBA transparency and support for up to 10 reference images. Product cutouts and clothing try-on are the main practical examples.

Setup requires three downloads: the diffusion model, a Qwen3-VL text encoder and the VAE. The walkthrough discusses BF16 and lower-memory options, including Int8 Convrot and a W4A8 text encoder. The speaker puts the combined Int8 diffusion model and text encoder at about 16 GB and says offloading allows use below that GPU memory capacity. Reported generation times reflect the presenter's setup, rather than a guarantee for other hardware.

The official workflow provides a starting point for local AI image generation. A separate custom Prompt Enhancer node uses Qwen's prompt rewriting models for text-to-image and editing prompts; the basic workflow does not require custom nodes.

Tests include technical diagrams, a trading terminal, clothing transfers and a room mockup using five references. The presenter finds that clothing details often survive edits, but blueprint dimensions conflict and room objects have unreliable scale. Adding a character sheet helps one camera rotation test, while room geometry still contains errors.