Qwen Image 2.1 in ComfyUI: setup and editing tests

Qwen Image 2.1 combines text-to-image generation and image editing in a single 7 billion parameter model, according to the presenter. This walkthrough explains its ComfyUI workflow and tests portrait detail, typography, object placement and targeted edits. The presenter lists Linux, an RTX 3060 with 12 GB VRAM and 48 GB system RAM as the test setup, rather than minimum requirements.

Setup uses the image model, a Qwen3VL 8B text encoder and the Qwen Image 2.1 VAE. The workflow supports native output up to 2K and as many as 10 reference images per edit. For text-to-image generation, the presenter enables custom sizing and disables image inputs. Editing can use a chosen resolution or inherit the first input image's dimensions. The RG3 group bypasser is omitted because the presenter reports a problem with the ComfyUI update used for these tests.

The presenter reports readable storefront lettering and mostly accurate longer text, with a punctuation error in the handwritten letter test. Clothing edits and compositions using several references largely preserve the original subject, though some details change. Transparency is less successful: generated RGBA images and background removal show dithering and noisy alpha masks that need cleanup before production use. The presenter says all tests ran locally in ComfyUI, making this a practical local AI workflow demonstration with specific output limitations.