Qwen Image 2.1 in ComfyUI: editing and upscaling tutorial

Nerdy Rodent demonstrates Qwen Image 2.1 in ComfyUI for text generation within images, image editing and reference-based composition. The tutorial uses the 7.26 GB INT8 model file, which the presenter reports gives the best performance on his RTX 3090. That file size is not a minimum VRAM requirement. He also says the model no longer has the open source licensing freedoms of earlier releases.

The workflow walkthrough explains positive and negative prompts, reference resizing and output dimensions. At CFG 1, the negative prompt has no effect. The presenter says the text encoding node's latent output is always 1024 by 1024, so he connects an empty latent to choose other sizes. Examples include a 4096 by 1024 image, though several results still have malformed fingers.

Two upscaling approaches show different tradeoffs. A six-step pass at 0.25 denoise keeps the example close to its original appearance. A prompted regeneration at 25 steps adds detail but also changes colors. An image-to-image example at 0.85 denoise changes the style while failing to add the requested banquet background.

For more precise local AI editing, the tutorial uses a manual mask to avoid changes outside the selected area. Later examples use style references, depth maps and pose input, then combine seven reference images into one scene. The node supports up to 10 references.