Seedream 4.5

Introduction

Seedream 4.5 renders a single positivePrompt at 2K to 4K resolution, holding fine detail and legible small text across the whole frame. It can read a short prompt and compose the scene for you, and it follows a long, layered prompt closely when you want to place every element yourself. The prompt budget runs to 2000 characters, room enough for a full scene breakdown.

That gives you two ways to work. Write a short prompt when you are exploring and want the model to make the compositional calls. Write a layered prompt when you have a specific image in mind and need the subject and the lighting placed deliberately.

A wide editorial photograph of an artisan coffee roastery interior in the morning, low sun raking through tall windows across burlap sacks stamped with origin stencils, a large matte-black drum roaster in the center, a chalkboard menu on the exposed brick wall reading "SINGLE ORIGIN" above a short handwritten price list, floor-to-ceiling shelves of glass jars with small printed labels, shallow depth of field on the roaster, warm tungsten light mixing with cool daylight, photorealistic, crisp legible signage

The roastery above came from a layered prompt that named the roaster, the raking light, the chalkboard copy, and the labeled jars. Seedream placed each one and kept the signage readable at full resolution. This guide covers both prompt styles, the layer order that hands you control, keeping rendered text legible, and the formats Seedream handles best.

Short prompts and layered prompts

Give Seedream a short prompt and it works in an interpretive mode, filling in composition and lighting on its own. Under about ten words, the model is making most of the aesthetic decisions for you.

a red fox in fresh snow at dawn

a neon ramen shop in the rain

Both prompts above are under ten words. Seedream chose the framing and the palette. Short prompts suit early exploration, when you want a direction rather than a finished result.

A layered prompt does the reverse. It spells out each part of the scene so the model has less to invent. The pair below uses the same subject: the first prompt is five words, the second names the headland, the light, the camera, and the mood.

a lighthouse on the coast

A weathered white stone lighthouse with a red lantern room standing on a rocky headland, waves breaking against black volcanic rocks below and a narrow gravel path climbing through windblown coastal grass, wide establishing shot from the foot of the path looking up, rust streaks on the gallery railing and a faint beam catching the haze, stormy blue-hour light broken by a thin band of warm orange on the horizon, shot at 35mm with deep focus, lonely and windswept atmosphere

The short prompt returns Seedream's reading of "a lighthouse on the coast". The layered prompt returns yours. Where the short version leaves the light and the framing to the model, the layered version names them, so the output tracks your intent.

[
  {
    "taskType": "imageInference",
    "taskUUID": "a1b2c3d4-e5f6-7890-abcd-ef1234567890",
    "model": "bytedance:seedream@4.5",
    "positivePrompt": "A weathered white stone lighthouse with a red lantern room standing on a rocky headland, waves breaking against black volcanic rocks below and a narrow gravel path climbing through windblown coastal grass, wide establishing shot from the foot of the path looking up, rust streaks on the gallery railing and a faint beam catching the haze, stormy blue-hour light broken by a thin band of warm orange on the horizon, shot at 35mm with deep focus, lonely and windswept atmosphere",
    "width": 2048,
    "height": 2048
  }
]

The prompt layers

A layered prompt is easier to control when you build it in a fixed order, from the most important element down. Seedream reads the entire string, but leading with the subject and adding context after it keeps the model from anchoring on a secondary detail.

1. Subject: who or what the image is about.
2. Environment: where it sits.
3. Framing: how the camera sees it.
4. Details: surface materials and identifying marks.
5. Lighting: the direction and color of the light.
6. Camera: the lens and depth of field.
7. Mood: the emotional read.

You rarely need all seven layers. A clean product shot might use only the subject and the lighting. A character scene might use every one. Add the layers the image actually calls for and leave the rest out.

Ordering and front-loading

Seedream weights the start of the prompt most heavily. Put the subject and its defining attributes in the first clause, ahead of the environment and the styling. A prompt that opens with a long stylistic preamble and reaches the subject late tends to render the subject weakly, because the model has already committed to the framing it read first.

The 2000-character budget is generous, so the real constraint is clause order, not length. Lead with what the image is of, then describe how it looks, and keep each layer in its own clause so the model can separate them.

Rendering legible text

Seedream renders text more reliably than most image models, and the 2K to 4K output is what keeps it sharp at small sizes. To fix the exact wording, wrap the literal text in quotation marks. Copy inside quotes is rendered as written. Words outside the quotes are read as scene description.

A photograph of a corner bakery storefront at dusk, a hand-painted sign above the door reading "GOLDEN CRUST BAKERY" in gold serif capitals, a small chalkboard propped in the window reading "Fresh Sourdough Daily", warm interior light spilling onto the wet cobblestone pavement, a striped awning and potted bay trees by the entrance, photorealistic, sharp and legible lettering

A vintage screen-print travel poster, bold condensed title text "VISIT THE DOLOMITES" arched across the top, a stylized illustration of jagged limestone peaks catching pink alpenglow at sunrise with a pine forest below, a single line at the bottom reading "NORTHERN ITALY", limited palette of warm coral, deep teal, and cream, flat halftone texture, clean legible typography, vertical format

The storefront sign reads "GOLDEN CRUST BAKERY" and the window board reads "Fresh Sourdough Daily" because both strings were quoted in the prompt. The poster keeps its title and its subtitle apart for the same reason: each line is quoted and given a position. For layouts with several text elements, quote each string and say where it goes rather than describing the copy in prose.

[
  {
    "taskType": "imageInference",
    "taskUUID": "b2c3d4e5-f6a7-8901-bcde-f23456789012",
    "model": "bytedance:seedream@4.5",
    "positivePrompt": "A vintage screen-print travel poster, bold condensed title text 'VISIT THE DOLOMITES' arched across the top, a stylized illustration of jagged limestone peaks catching pink alpenglow at sunrise with a pine forest below, a single line at the bottom reading 'NORTHERN ITALY', limited palette of warm coral, deep teal, and cream, flat halftone texture, clean legible typography, vertical format",
    "width": 1664,
    "height": 2496
  }
]

Working by format

Seedream produces both photographic and illustrated work from the same request. The prompt language sets the format, so naming the medium early steers the model before it commits to a look.

Photography

Brief it like a photographer. Name the lens and the light source, and Seedream renders the spatial cues accurately.

A studio portrait of an older fisherman with a deeply weathered face and short white stubble, wearing a worn navy wool sweater, soft north-window light from the left modeling the skin, shallow depth of field at 85mm, muted cool palette, calm and direct gaze, editorial photography

A product photograph of a faceted glass perfume bottle with an amber-to-clear gradient and a brushed gold cap, standing on a wet black slate surface scattered with water droplets, a single soft key light from the upper right with a controlled reflection below, minimal near-black background, premium beauty product photography

Illustration and design

For illustration, name the medium and the mark-making, such as gouache or flat vector. For posters and layouts, describe the type hierarchy and quote the copy so the design reads as intended.

A storybook illustration of a small wooden sailboat crossing a calm sea under a thin crescent moon, painted in gouache with visible brushstrokes and soft paper texture, a palette of dusty teal, deep navy, and warm cream, gentle hand-drawn style, a scattering of stars reflected on the water

A modern minimalist concert poster, large sans-serif title "MIDNIGHT ECHO" centered in the upper half, a smaller subtitle below reading "LIVE AT THE FORUM · OCT 12", abstract concentric sound-wave line graphics radiating from the center, high-contrast palette of near-black and electric blue with a single warm accent, clean print layout with generous margins, crisp legible type, vertical format

Tips

1. Lead with the subject. Seedream weights the front of the prompt most, so the first clause should name what the image is of. Styling and mood come after.

2. Quote text you want rendered. Anything in quotation marks is treated as literal copy. Without quotes, the model reads the words as scene description and may paraphrase them.

3. Keep text strings short. Headlines and signs render cleanly. Long paragraphs degrade. For dense copy, composite separate text elements.

4. Match the dimensions to the shot. Seedream renders between roughly 2560 × 1440 and 4096 × 4096. Use 2048 × 2048 for square, 1664 × 2496 for a portrait poster, and 2560 × 1440 for a wide scene. The model page lists every supported size.

5. Build up one layer at a time. Start with a short prompt, read the model's interpretation, then add the layers that pull it toward your intent. This is faster than writing all seven layers blind.

6. Reach for 4K when the detail carries the image. Dense scenes and small text benefit from the higher resolution. For quick iteration, stay at 2K and scale up once the composition is right.