Qwen Image 2.1

The clearest strength on paper is transparent asset generation. When background is set to transparent, the output PNG or WebP carries a genuine alpha channel rather than a background color that merely looks removed — which matters if the image is going into a design tool, a video overlay, or an e-commerce listing where the transparency itself needs to survive further editing. The tradeoff is that transparency can't be paired with JPEG output, and the prompt should describe only the subject; mentioning a background or scene in the prompt can push the model toward rendering a non-transparent result even with the toggle on.

Sizing is exact rather than approximate. Each of the nine supported aspect ratios maps to a fixed pixel size at each resolution tier — for example 3:2 renders at 1248×832 (1K) or 2528×1696 (2K), and 9:16 at 768×1376 or 1536×2752. Because text-to-image defaults to 3:2 and every ratio has a defined 1K and 2K size, you know the output dimensions before generating, which is useful when the image has to fit a predefined slot (a banner, a print size, a video frame) rather than getting cropped after the fact.

Prompt language isn't restricted to English — the model accepts prompts up to 5,000 characters in any language. Combined with enhance_prompt, which expands a short prompt into a fuller scene description automatically, this makes the text-to-image tab usable for non-English prompts without a separate translation step, though the quality of that expansion for any given language isn't something VdoBloom benchmarks or claims a specific result for.

The Image Edit tab takes 1 to 10 reference images (JPEG, PNG, or WebP, each up to 30 MB and 25 MP) and reads them in the order you provide them — the prompt refers to them by position, as in "the first image" or "the second image." That ordering is worth getting right, since it's how the model knows which reference each instruction applies to when you're combining people, objects, or styles from multiple sources into one output.

Fidelity to any single reference drops as you add more of them, so staying within four images is the practical ceiling when you need a reference to be closely preserved rather than loosely referenced. When combining several people or objects in one composition, landscape aspect ratios (3:2 or 16:9) tend to give the composition more room to work with than portrait ratios. Aspect ratio for image-to-image defaults to auto, meaning it follows the first reference image unless you override it.

Passing a mask_url — a black-and-white mask image — switches the whole request into local-edit (inpainting) mode: white marks the area to change, black marks what stays untouched. That mode comes with hard constraints: exactly one reference image is allowed, the mask must match that image's aspect ratio, it can't be combined with transparent background, and the output ratio follows the reference image regardless of any aspect ratio setting — prompt rewriting is also ignored in this mode, since the mask is doing the work of scoping the edit rather than the prompt describing a whole new scene.

enhance_prompt is on by default on both tabs. It takes your prompt and rewrites it into a fuller scene description before the image is generated — useful for short or sparse prompts, but worth turning off if you've already written a precise, fully-specified prompt and don't want it altered before generation, or if you're using image-to-image with a mask (where prompt rewriting is ignored anyway).

seed controls reproducibility. Every generation returns the seed it used, and reusing that seed with the same prompt and settings reproduces the same image — which is the mechanism to use if you land on a result you like and want to regenerate it with a small prompt tweak while keeping everything else about the composition stable, rather than rolling the dice again.

Output format defaults to PNG, with WebP and JPEG as alternatives. The format choice interacts directly with the background setting: transparency requires PNG or WebP, since JPEG has no alpha channel, so if transparent background is the goal, JPEG isn't an option regardless of what's selected elsewhere.