AI background remover: a real transparent PNG, not a white fill

What it does: takes one uploaded photo, reads a text prompt describing the subject, and returns that subject on a transparent alpha background as a PNG or WebP. It is an image-to-image edit, not a generate-from-scratch job — the output keeps the subject from your photo rather than reimagining it, since the underlying request is Qwen Image 2.1's image-to-image mode with your uploaded file as the single reference image.

What it does not do: it is not a dedicated, single-click "cutout" tool with its own specialized segmentation model — there is no separate background-removal AI in VdoBloom's codebase. It is one setting (background=transparent) on a general-purpose image editor, which means the quality of the cutout depends on the same prompt-following behavior as any other edit on that model, not on a purpose-built matting algorithm. It also cannot be combined with Qwen 2.1's mask-based local-edit mode in the same request — if you need to both mask out an object and make the background transparent, that has to be two separate passes, not one.

A related but different capability exists elsewhere on the platform: most other image-edit models (the default Nano Banana Edit included) will follow a plain instruction like "remove the background" and return something plausible-looking. That is real, but unverified for true transparency in this codebase — those models don't expose a background/transparent parameter the way Qwen Image 2.1 does, so there's no code-level guarantee they're producing an actual alpha channel rather than a flattened, opaque fill. If a genuinely transparent file is the goal — for compositing, for a product listing platform that requires alpha PNGs, for a sticker or logo asset — Qwen Image 2.1's transparent toggle is the verified path on VdoBloom; treat a generic "remove background" prompt on another model as a look-alike, not a substitute.

Qwen Image 2.1 is Alibaba's Qwen 2.1-generation image model, exposed on VdoBloom through KIE's jobs API as two model ids — a text-to-image tool and this image-to-image editor. The editor reads your uploaded photo as a reference, your prompt as an instruction, and a set of parameters: resolution (1K or 2K), output format (PNG, WebP or JPEG), and the background switch (opaque or transparent). When background is set to transparent, the server enforces one hard rule before anything is sent to the model — output format must be PNG or WebP, because a transparent result in JPEG is a contradiction (JPEG has no alpha channel at all), so a JPEG selection combined with the transparent toggle is refused up front rather than silently producing an opaque file and calling it done.

The prompt still matters even though the job is "just" removing a background. Because this is a general image-to-image edit and not a specialized matting model, whatever you describe is what the model tries to render — describing the subject precisely (what it is, its pose, its material) gives the model something concrete to preserve, while describing a background or scene gives it something to paint, which works against the transparency you asked for. The enhance_prompt setting, on by default, rewrites your prompt into a fuller scene description before generating; for this specific job it's usually worth turning off, since an automatic rewrite could add scene detail you didn't intend and never see before it's applied.

Every parameter here — resolution, aspect ratio, output format, transparency — is billed the same way regardless of the combination you choose: by resolution tier alone. That's different from some other models on the platform where extra features (higher reference counts, larger canvases) raise the price. A transparent 1K cutout and an opaque 1K edit both cost 3 credits; the transparency toggle is free to flip.

Product cutouts for marketplace listings and D2C storefronts, where a plain alpha PNG can be dropped onto a template background without a visible seam — useful when you already have a photo of the product and just need the surrounding clutter gone rather than a fully staged new shot.

Sticker-style and logo assets, where the whole point of the file is that the edges are transparent rather than boxed in by a colored background — a chat sticker, a watermark-free logo layer, or an icon meant to sit over other artwork.

Portrait and headshot compositing, where a person needs to be lifted cleanly off their original background so they can be placed over a different one in a separate design tool, video editor, or presentation slide.

Any asset headed into further editing where the removed area needs to stay genuinely empty — video overlays, motion graphics layers, or any pipeline where a flattened white or gray background would show up as a visible rectangle once composited.

The numbers below come directly from the model's validation code and VdoBloom's credit system — nothing here is an estimate.

Write the prompt as a description of the subject alone — "a red leather handbag, front-on" rather than "a red leather handbag on a marble table" — since anything you name as a surrounding element gives the model a reason to render it, working against the transparent result you want.

Turn off enhance_prompt for this job specifically. It's a useful feature for scene generation, where an automatic rewrite fleshes out a short idea, but for a background-removal edit it's an unnecessary variable — you want exactly what you typed, not an expanded version of it.

Choose PNG when the subject has hard edges or flat colors (logos, text, product packaging) and WebP when file size matters more and the subject has soft gradients or photographic detail — both carry the alpha channel correctly, the choice is about file size and compression style, not about whether transparency works.

Use 2K when the output is headed somewhere that will scale or crop it further (a print asset, a large hero banner); 1K is enough for most web and social placements and costs less.

If the result keeps a faint background instead of a clean cutout, the most common cause is a prompt that named or implied a setting — rewrite it to name only the subject and try again before assuming the model can't do the job.