Qwen-Image-2.1 supports native 2K image generation, so I gave it a try as an upscaling/detail-enhancement method without using any additional upscaler nodes or LoRAs.

If you see a “missing node” warning on the Math Expression nodes, you can replace them with ComfyUI’s native Math Expression node and copy the same expressions.

The Math Expression nodes automatically calculate the output resolution while preserving the input aspect ratio. The calculation uses approximately 2048×2048 worth of total pixels (~4.2 MP) as the maximum 2K pixel budget, rather than limiting each side to 2048 px. For example, a 16:9 image can end up around 2730×1536, which is close to Qwen’s officially recommended 2752×1536 16:9 resolution.

You do not have to use these Math Expression nodes. You can also enter your preferred width and height manually in the Empty Latent Image node.

What matters most here is the prompt. The prompt included in the workflow is intentionally generic and should work for many types of images. However, if you get excessive grain, noise, or other artifacts, try adjusting the prompt.

For example, the prompt below is longer and more detailed than the one I use in the workflow. It can improve portraits and bring out more fine detail, but on some images it can also introduce noticeably more artifacts:

Enhance this image to high resolution with rich fine details. Sharpen textures and surfaces, add natural microdetails, and improve clarity and definition across the image while preserving the original composition, lighting, atmosphere, and visual style. The input may be an illustration or a real photograph; preserve the original style.