Qwen-Image-Layered Guide to Creating Editable RGBA Image Layers

Qwen-Image-Layered separates an existing image into editable RGBA layers for moving, resizing, or recoloring individual components. Alibaba's Qwen team publishes Apache 2.0 weights and a QwenImageLayeredPipeline example that exports PNG files. Choose the layer count in the pipeline, then inspect the exported components before editing them. Model card Official repository

What are the key facts about Qwen-Image-Layered?

RGBA includes an alpha channel, which carries transparency alongside color. The model card's examples use that representation to separate parts of a composition and change a chosen layer while leaving the other layer assets alone. Model card

What can you edit with Qwen-Image-Layered?

The documented operations include moving, resizing, recoloring, and removing components after decomposition. Qwen also demonstrates editing an isolated layer with Qwen-Image-Edit, including replacing a subject or changing text. These examples make the model useful to investigate when a design needs several controlled revisions around the same composition. Model card

Consider a promotional illustration containing a character, a sign, and a background. A useful evaluation asks whether the decomposition gives you the components needed for the planned revisions. If the sign and character remain inseparable, inspect a finer decomposition before committing to a sequence of changes.

The model supports a variable layer count and recursive decomposition: a layer can be submitted again for further separation. Qwen illustrates different decompositions of the same image. Treat that flexibility as a way to explore useful structure, rather than assuming the most detailed decomposition is always the easiest to edit. Model card

For a repeatable asset workflow, keep the original image, untouched generated layers, and edited derivatives in separate folders. Give each accepted layer a descriptive filename and record its stacking order. This suggested organization makes it easier to undo an unsuccessful edit and explain which component changed.

What are the limits of Qwen-Image-Layered decomposition?

The documented workflow exports raster RGBA layers as PNG files. Check whether those exported components provide the separation your intended revisions need, especially around text and overlapping objects. Model card

Layer isolation helps preserve assets that you do not edit, but that is a different question from whether the initial decomposition accurately represents the source. Review the recomposed output before editing. Pay particular attention to object boundaries, fine details, and the areas revealed when moving a foreground component.

The quick start provides resolution buckets of 640 and 1024 and recommends 640 for this version. Those are documented pipeline settings; they should not be turned into a promise that every input keeps its original dimensions or that arbitrary layer counts have the same memory cost. Model card

A universal VRAM minimum and guaranteed decomposition time are not published in the model card. Because the provided example runs a large checkpoint on CUDA, plan a small representative trial before deciding that a given workstation can handle a full asset collection. Model card Weight metadata

How do you export RGBA layers with Qwen-Image-Layered?

The model card asks for Transformers 4.51.3 or later, installs Diffusers from its Git repository, and includes python-pptx in its setup commands. The direct PNG-export example below uses Pillow, PyTorch, and the documented pipeline; it does not create a presentation file. Model card

Start with your own composition.png, then run this shortened version of the official example in the configured environment:

import torch
from PIL import Image
from diffusers import QwenImageLayeredPipeline

pipe = QwenImageLayeredPipeline.from_pretrained("Qwen/Qwen-Image-Layered")
pipe = pipe.to("cuda", torch.bfloat16)
with torch.inference_mode():
    layers = pipe(
        image=Image.open("composition.png").convert("RGBA"),
        generator=torch.Generator(device="cuda").manual_seed(777),
        negative_prompt=" ", true_cfg_scale=4.0,
        num_inference_steps=50, layers=4, resolution=640,
        cfg_normalize=True, use_en_prompt=True,
    ).images[0]
for index, layer in enumerate(layers):
    layer.save(f"layer-{index}.png")

The sample requests a decomposition with layers=4 and saves each returned image. Keep every output while inspecting the result; do not discard files based only on their numeric names. The card explains that the pipeline can create an automatic caption when the user does not provide one. Model card

Next, open the exported PNGs in an image editor that can display transparency. Recompose the layer stack, compare it with the input, and make a small change such as moving one object. This is a proposed acceptance check for your assets, not an additional model capability claim.

If an isolated object needs a semantic change, the Qwen-Image-Edit guide covers instruction-based editing. For work involving several character references, the sibling Qwen-Image-Edit-2511 guide addresses a different editing task.

How does Qwen-Image-Layered compare with image editing?

These models can serve different stages of a project without being interchangeable checkpoints. The ComfyUI complete guide explains workflow organization if you want to manage an image process visually; the concrete local route shown here is the model card's Python pipeline.

What should you know before using Qwen-Image-Layered?

What files does Qwen-Image-Layered export?

The Qwen-Image-Layered Python example saves each returned RGBA layer as a PNG file. Inspect the exported layers in an editor that displays transparency before moving or recoloring individual components. Model card

Can I choose the number of Qwen-Image-Layered layers?

Qwen-Image-Layered exposes a layers setting, and Qwen demonstrates different layer counts for the same image. A generated layer can also be submitted for further decomposition. Model card

Can I edit text on a Qwen-Image-Layered layer?

Qwen demonstrates using Qwen-Image-Edit to revise text on a layer produced by Qwen-Image-Layered. Separate the image first, edit the selected layer, then inspect the lettering and boundaries before recomposing the image. Model card

Can Qwen-Image-Layered run locally?

Qwen-Image-Layered has downloadable Apache 2.0 weights and a CUDA-based Diffusers example. The model card also links a demonstration for exploring decomposition before configuring a local environment. Model card

Sources

• Qwen-Image-Layered model card
• Qwen-Image official repository and release history
• Qwen-Image-Layered published weight metadata
• Qwen-Image-Edit model card
• Qwen-Image model card

Related guides on PromptZone

• Local LLMs 2026: Run Llama, Mistral, Qwen on Your Hardware