Seedream 5.0

How Seedream 5.0 reads a prompt

Seedream 5.0 comes in two variants. Lite is the fast variant, and ByteDance describes it as a unified multimodal model with deep thinking and online search, built to infer the intent behind a request and to visualize current topics. Pro is the flagship: ByteDance highlights text-dense images such as infographics, edits that follow spatial annotations and sketches, photographic lighting and skin texture, and native support for about a dozen widely used languages.

Both variants read natural language, and both treat each clause of a prompt as an instruction. A prompt that separates the format, the subject, the composition, the light, any text and the style tells the model what to do in every layer. Leave a layer out and the model fills it with its own defaults. Runware's guide for Pro recommends keeping prompts under about 600 English words, because detail starts to drop out in longer prompts.

A structure that works

Write the layers in this order, one or two sentences each:

[Format and orientation]. [Subject and its details].
[Where things sit in the frame]. [Light and camera].
[Exact text in quotes, with style and position]. [Style].

Lead with the format. "Portrait 3:4 magazine cover" or "square e-commerce packshot" tells the model what kind of image it is composing before it reads the details, so it reserves room for a masthead or centers the product. In Endless the size setting picks a resolution tier rather than an aspect ratio, so state the orientation or ratio in the prompt.

Be specific about materials and light. "Matte terracotta ceramic" and "soft overhead studio light with a gentle contact shadow" are far more specific than "nice lighting". Name the lens or the angle when the shot needs it: "shot on a 35mm lens at knee height", "seen from directly above".

Text in images

Seedream 5.0 renders text when the prompt names it. There is no separate text field.

• Put the exact copy in quotes. A paraphrase such as "a French title" produces an invented string.
• Name the script and the type style together: "tall condensed serif capitals", "hand-lettered script", "vertical mincho serif".
• Say where each line sits: "across the top third", "down the right side", "in small capitals at the bottom".
• Give each block of copy its own sentence with its own size, color and position. Multi-script layouts, such as a Japanese title with an English footer, work best when each script is described separately.
• Keep text short. Titles, subtitles, prices and labels render reliably; a paragraph of body copy is where typos appear. Small print and legal text are better added in a layout tool afterward.

Runware's testing of Pro lists 15 languages it renders natively, including English, Spanish, Portuguese, French, German, Japanese, Korean, Arabic and Thai. Other languages still generate, but text quality is weaker. Have a native speaker check non-Latin copy before it ships.

References and editing

Attach images and describe the change, and Seedream 5.0 edits or combines them. Endless accepts up to 10 reference images.

• Name what must not change. End an edit with a clause such as "Keep the exact silhouette, camera angle, backdrop and lighting." Without it, anything the prompt does not mention can drift.
• One edit, clearly stated. "Change the jacket to mustard yellow corduroy" is easier to follow than a list of loosely related changes. Make further changes in a new round.
• Refer to references by what they show. "The ceramic mug" and "the steel grinder" are more reliable than "the first image", especially when you reorder or swap references.
• Use references for identity, prose for the rest. A clean packshot preserves a product's shape, materials and logo more reliably than any description. Plain-background product shots combine more cleanly than busy lifestyle photos.
• Merging references. Say where each element goes and how they relate: "mug in front, grinder behind it on the left". The references set each object's look; the prompt sets the layout and the light.

Pro is described by ByteDance as responding to spatial annotations and sketch guidance, so a rough sketch of a layout can serve as a reference for a finished render.

Settings that matter

• Mode. Lite is the fast variant and reaches 3K. Pro is the flagship and goes up to 2K. Draft and explore with Lite, then move to Pro for text-dense layouts or edits that combine several specific references.
• Image size. Lite offers 2K or 3K; Pro offers 1K or 2K. For dense text, such as footer lines or small subtitles, use the larger tier available, since small type renders more cleanly at more pixels.
• Reference images. Up to 10. Leave the field empty to generate from the prompt alone.

Common failure modes and fixes

• Unwanted changes after an edit. Add or strengthen the "keep" clause, naming each element: subject, pose, background, light.
• Text in the wrong place or hidden behind the subject. Give the text a position and describe the empty area it sits on, such as "a plain cream wall across the top 40% of the frame".
• Garbled or invented words. Quote the copy exactly, shorten it, and split long lines into separate blocks.
• Generic results. Add the missing layer, usually the format, the light or the camera.
• Details dropped from a long prompt. Make vague clauses specific instead of adding more; move secondary details into a follow-up edit.