Qwen-Image-Layered is free to use: the Qwen team released the model under the Apache 2.0 licence, which allows commercial use. What is not free is the GPU it needs, so running it for free means either the official online demo, which has a daily GPU quota, or your own graphics card.
Disclosure: this page is published by LayerGrab, a hosted tool that splits images into layers with a different model. Every limit below comes from the source linked at the bottom, checked on 2026-10-06. For what the model is and how it works, read What is Qwen-Image-Layered?
Last updated 2026-10-06
Runs on shared ZeroGPU hardware with a daily quota: 2 minutes of GPU time without signing in, 5 minutes with a free Hugging Face account, 40 minutes with PRO. The quota resets 24 hours after your first run, and you may queue at busy times.
Good for: Trying the model on a few images
Free software; your hardware and power
Supported natively. ComfyUI’s guide measured a 45 GB memory peak and 120 seconds for one 1024 px image on a 96 GB RTX Pro 6000; RTX 4090 owners report it fills almost all 24 GB. A lower-memory fp8 file is offered.
Good for: Batches, and images that must stay on your machine
Free software; your hardware
The README’s QwenImageLayeredPipeline on a CUDA GPU, with the same memory needs as ComfyUI.
Good for: Building layer decomposition into your own code
Hosted API
Not offered by the Qwen team
The model card says no Hugging Face inference provider deploys it. Hosted services that run it online have their own plans.
Good for: Using it without your own GPU
The free demo is enough to see what the model does, not to work with it every day. Five minutes of GPU time a day on a free account comes to a few images at 1024 px, and fewer when the queue is long.
To stretch it: run at 640 px, which the README recommends for this version anyway, and ask for only as many layers as you need, since you choose the count before each run. If you keep hitting the limit, your choices are a Hugging Face PRO plan, a GPU of your own, or a hosted tool.
LayerGrab splits an image into separate transparent layers online, with a different model, in about a minute. You can explore every feature on the sample images without an account. New accounts in supported regions get 30 free credits, enough for a split of up to 3 layers.
It differs from the open model in what you control. You do not set a layer count; the model decides from the picture, and you can name the one element you want, such as “only the price tag”. Each layer comes back with an English name, and you can download PNGs, a ZIP or a layered PSD.
You have a GPU with enough memory, need dozens of runs a day at no cost per image, or your images must never leave your machine. LayerGrab cannot give you any of those.
You have no suitable GPU, the demo quota runs out, you need to pick one element out of a picture, or you want the layers placed in Photoshop, Figma or Sketch.
Yes. The Apache 2.0 licence lets you use, modify and ship the model commercially, as long as you keep the licence notice. The licence covers the model only: you still need the rights to the images you split.
A free Hugging Face account gets 5 minutes of shared GPU time a day, and visitors who do not sign in get 2. ComfyUI’s guide measured about two minutes for one 1024 px image on a full RTX Pro 6000, and ZeroGPU’s default is half of that card, so expect a handful of images a day, not an afternoon of work. The README recommends 640 px for this version, which uses less of the quota.
The Qwen team does not publish a minimum. ComfyUI offers an fp8 file that needs less memory, and the model card lists community quantized versions that are smaller still, at some cost in quality. We have not tested them, so treat any VRAM figure you read as a starting point.
Not from the Qwen team. Its model card says no Hugging Face inference provider deploys the model. Hosted services that offer it online run it on their own hardware, each with its own plans.
No. LayerGrab runs a different hosted model. It is on this page as a free-to-try alternative for people who do not have a GPU, not as a way to run the Qwen model.