Prompt design
Qwen-Image-2.1 prompt guide for transparent PNGs
A transparent result is easier to control when the prompt describes a subject, not a scene. This guide covers the required RGBA opening, a repeatable prompt formula, and the wording that most often creates an unwanted background.
The transparent opening that matters
Begin with the exact sentence: “This is an RGBA image with transparency.” It establishes the output format before the model encounters any scene language. The next sentence should state what is transparent: the background, the area around the subject, or every pixel that is not part of the subject.
Alpha is easier to preserve for compact subjects with clear silhouettes. Products, icons, stickers, isolated characters, and individual objects are natural candidates. A scene containing several people, architecture, smoke, and reflected surfaces can still be generated, but the final edge often needs more manual inspection.
Minimum transparency instruction
This is an RGBA image with transparency. Place the complete subject on a fully transparent background. Do not add a floor, backdrop, border, text, or cast shadow.
A subject-only prompt formula
After the RGBA sentence, describe the output in this order: subject, view or pose, materials and colors, lighting on the subject, edge requirements, and exclusions. Framing words such as “full body,” “complete object,” and “small clean margin” are useful. Environment words such as “studio room,” “on a table,” or “urban street” work against transparency.
Product cutout
This is an RGBA image with transparency. A brushed stainless steel water bottle centered upright, three-quarter product view, crisp machined edges, soft studio highlights, no cast shadow, no surface, no backdrop, no text.
Character asset
This is an RGBA image with transparency. A friendly cartoon fox standing in a neutral pose, full body visible, clean silhouette, soft fur edges, flat game-asset lighting, no floor, no scenery, no border.
Photo isolation
This is an RGBA image with transparency. Isolate the woman in the source image and preserve her hairstyle, clothing colors, sunglasses, and natural proportions. Keep loose hair edges, remove the environment, do not add a new background or shadow.
The Pellucen generator can rewrite a short brief into this structure, then lets you edit the result before creating the image. Open the transparent background generator or read the ComfyUI workflow guide for local execution.
Fixing common transparency prompt failures
A checkerboard pattern in the prompt is a bad shortcut. It describes pixels, not transparency, and can cause the model to paint an actual checkerboard background. Say “fully transparent background” and use the output settings to request PNG.
Avoid asking for a shadow “under” the subject when the shadow must disappear. An object can keep internal contact shading, but a cast shadow needs a receiving surface and usually makes the image opaque. For hair, fur, glass, or smoke, ask for preserved soft edges and inspect the final alpha channel at 100% zoom.
If the model adds a frame or label, move those words to the exclusion portion of the prompt. If it leaves part of the original environment around an uploaded subject, name the specific object to remove and the details to preserve. For file-level guidance, continue to the transparent PNG maker.
Should every Qwen-Image-2.1 prompt start with the RGBA sentence?
For transparent output, yes. Starting with the exact RGBA sentence makes the desired output format unambiguous before the subject description begins.
Can I mention a white background and make it transparent later?
Avoid that approach. A white background asks the model to generate an opaque scene. Request transparency directly, then use the alpha channel to inspect and refine the result.
How long should a Qwen-Image-2.1 prompt be?
Use enough detail to define the subject, framing, material, and edge behavior, but remove unrelated scene descriptions. A focused prompt of two to five sentences is usually easier to control than a long story.