Skip to content
CozyGPU

Qwen Image 2.1

Text-to-image and reference-based image editing with up to ten input images.

Modes
Text to image · Image edit
GPU class
48 GB Ada Pro (falls back to 32 GB Ada Pro)
RunPod pool price
$1.58 to $1.75/h
Typical cost per run
about $0.01 per image
01

What it is

Qwen Image 2.1 is an open-weight image model that is as comfortable editing as it is generating. In edit mode you attach one to ten reference images and address them in the prompt as <image1>, <image2> and so on, which makes product swaps, outfit changes, and style transfers a sentence rather than a mask.

It is slower and heavier than Z-Image Turbo, so CozyGPU schedules it on a 48 GB class GPU. Expect about 25 seconds a picture when the worker is warm.

Weights on Hugging Face: Comfy-Org/Qwen-Image-2.1
02

Modes

  • Text to imagePrompt in, image out. Default 1024 x 1024.
  • Image editOne to ten reference images, referred to in the prompt as <image1> … <image10>.
03

Settings exposed

  • SizeAspect presets; default 1024 x 1024 (text to image). Edits follow the reference.
  • Steps10 to 50, default 25.
  • SeedRandom, or fixed.
  • Reference imagesUp to ten in edit mode. Files over 7 MB go through a network volume in your account.
  • Guidance and samplerFixed by the workflow (cfg 1, euler / simple; with a LoRA chained, er_sde / beta at cfg 3.5).
  • LoRAsBuilt-in realism LoRAs plus your own, each with a strength slider and trigger words.
  • Custom LoRAsSupported. Two built-in realism LoRAs (Lenovo UltraReal, SamsungCam UltraReal) ship with the worker; your own .safetensors files load by drag and drop and work in both text to image and image edit, up to three per job.

Other models