Seedream 5.0 Pro Adds Sketch-Guided AI Image Editing

Seedream 5.0 Pro is positioned as a way to turn AI images from one-shot renders into material that can be directed, revised, and reused. The practical problem is familiar: an image can land the mood and composition, but a misplaced object, a headline that needs moving, or copy that needs localization can force a complete regeneration.

ElevenLabs’ answer combines more precise inputs with more usable outputs. Users can mark up an existing image to direct an edit, generate dense layouts with comparatively readable text, and create text directly in 14 languages. The most consequential output feature—separating a composition into transparent layers—is due shortly after launch rather than available in the initial release.

Pro is described as a larger model built on the same foundation as Seedream 5.0 Lite. Lite introduced a dedicated reasoning layer; Pro’s distinction is control over the final image. For design work, the question is not simply whether a model can create an attractive composition, but whether that composition can be adjusted without starting over.

At launch, Seedream 5.0 Pro generates at 2K. ElevenLabs says 4K is not yet available.

Sketch editing is the clearest example of input-side control. A user can draw a rough shape where an object should appear, scribble over the thing to remove, or annotate an image before upload. Paired with a short instruction, those marks are interpreted as directions rather than incorporated literally into the result.

The advantage is that location, scale, rough shape, and arrangement no longer need to be translated entirely into prompt language. A crude markup can carry the instruction directly: put something here, remove this, follow this layout.

A living-room example begins with rough drawings for a floor lamp and a large houseplant. Given the instruction, “Replace the sketch of the lamp and the houseplant with real objects,” Seedream returns a room with a modern lamp and leafy plant in the indicated positions. The sketches function as placement constraints, while the generated objects are made to fit the scene.

Other demonstrations extend the same interaction beyond simple insertion or removal. A hand-drawn jazz-poster draft becomes a Bauhaus-style concert announcement; a floor plan becomes a photorealistic living room following the indicated furniture layout; a curved line drawing becomes an earthship house with the same shape and window configuration. A four-panel stick-figure storyboard is rendered as polished 3D-animation frames. In each case, the drawing supplies structural information that would be cumbersome to specify precisely in prose.

Users can work from a prompt alone or upload up to 10 reference images. Marking up a reference image before uploading it is part of the intended workflow: the visual instruction and written prompt work together.

The four new controls operate at different stages. Sketch editing directs an edit through the input image. High-density generation creates a composed page, diagram, or information-rich graphic. Layer separation is intended to make the resulting composition editable after generation.

For dense layouts, Seedream 5.0 Pro is presented generating ecommerce homepages, technical exploded views, and annotated infographics. A pollination graphic includes a title, four numbered stages, biological labels, explanatory copy, and a dotted flight path. Camera and drone renders pair separated components with leader lines and technical names. The point is not isolated text rendering but the ability to arrange text, images, labels, and hierarchy into a coherent visual structure in one generation.

ElevenLabs says small-text performance is “dramatically better,” while acknowledging it is not flawless. It describes the output as usable for infographics, labels, and even distant street signs that need to remain readable within a generated scene.

Native multilingual generation extends that proposition beyond English-language layouts. Seedream 5.0 Pro is said to generate text directly in the target language rather than create copy in one language and translate it onto a completed design afterward. ElevenLabs specifically cites Arabic’s right-to-left layout and Thai tone marks, as well as target-language letterforms, typographic conventions, and local aesthetics.

The examples test whether a campaign design can remain intact through localization. A cycling advertisement is rendered in German, Korean, Arabic, Thai, Spanish, and Swedish while retaining its underlying visual composition. A book cover appears in Portuguese, Russian, Arabic, and Thai; a digital storefront billboard appears in Polish, Arabic, and Japanese. The intended use is not simply translating words, but changing text while preserving a design system.

A generated image is ordinarily a flat picture: if the layout works but the title needs to move, changing it can require regenerating the entire image. Seedream 5.0 Pro’s layer-separation feature is designed to split a generation into a background plus separately exported elements, each with a transparent background.

Users specify what they want extracted, from a single line of text or object to a group of graphics. ElevenLabs gives a range of two to 20 layers from one image. The Wufkin dog-toy poster illustrates the intended request structure: a blurred sunny-park background, a woman and golden retriever in the midground, a dog toy in the foreground, a wordmark, and dedicated tagline space. The displayed result breaks the poster into distinct components.

Each component returns as its own transparent file, according to ElevenLabs. A creator can rearrange the composition in another design application, or feed one component back as a reference image and generate something new around it. A successful generation can therefore yield multiple reusable assets rather than remain a single terminal image.

This capability is not part of the initial rollout. Multi-layer separation is due slightly after Seedream 5.0 Pro’s release.

Seedream 5.0 Pro is available in Eleven Creative through the Image & Video area. Select the Seedream 5.0 Pro model, enter a prompt, and optionally add reference images. The interface permits up to 10 references, including a marked-up image for sketch-directed editing.

The practical value of the release lies in more explicit visual direction and in attempting layouts or localized graphics that may otherwise demand repeated reruns. ElevenLabs’ view is that clearer control should reduce regeneration and, with it, wasted credits.

The limits are equally specific: output is currently 2K, not 4K, and layer separation is the feature to watch for after launch rather than one to assume is already available.