Seedream 4.5 Prompt Guide: Examples That Work

Short answer: write Seedream 4.5 prompts as one or two plain sentences that name the subject, what it is doing, where it is, and how it should look, then put every word you want printed in the image inside double quotes. ByteDance's own guide says Seedream 4.5 needs less description than older versions, so short, precise prompts usually beat long lists of adjectives. For edits, say exactly what changes and what must stay the same.

This guide turns ByteDance's official advice into a reusable formula, 14 prompts you can copy, a list of common mistakes, and a working request for running Seedream 4.5 on APIMart. All prompt examples are in English because ByteDance documents Chinese and English as the supported prompt languages for Seedream 4.5.

Key Takeaways

• Seedream 4.5 handles text-to-image, image editing, reference-based generation and multi-image output in one model, according to the BytePlus prompt guide.
• The official API accepts up to 14 reference images, and reference images plus generated images cannot exceed 15 per request (BytePlus API reference).
• Output is 2K or 4K, with a pixel range from 2560x1440 to 4096x4096 and aspect ratios between 1:16 and 16:1 on the official API.
• ByteDance recommends keeping prompts under 600 English words (or 300 Chinese characters), because very long prompts make the model drop details.
• Text that must appear in the image belongs in double quotes, for example the title "SUMMER SALE".
• On APIMart, Seedream 4.5 offers 11 aspect-ratio presets plus auto, 2K or 4K output, and batch generation when you pass at least one reference image.

Seedream 4.5 at a Glance

The ByteDance Seed product page describes the 4.5 release as an all-round upgrade that better identifies the main subjects in multi-image editing, keeps reference details intact, and improves typography and dense text rendering. Those three strengths shape how you should prompt it.

When Was Seedream 4.5 Released?

ByteDance Seed released Seedream 4.5 in December 2025 as the successor to Seedream 4.0 (September 2025). The official model ID, seedream-4-5-251128, points to a late-November 2025 build, and API platforms added it through December 2025. It was followed by Seedream 5.0 Lite in February 2026 and Seedream 5.0 Pro in July 2026.

The Seedream 4.5 Prompt Formula

ByteDance's first rule is to describe the scene in coherent natural language: subject, action and environment, then style, color, lighting and composition when the look matters. We add one more slot for text, since that is where Seedream 4.5 is strongest.

The seven slots

You do not need every slot every time. A portrait might skip text, and a poster might skip action. The order matters less than the clarity: write it the way you would brief a photographer or designer.

Sentences, not tag lists

The official guide gives a clear contrast. "A girl in a lavish dress walking under a parasol along a tree-lined path, in the style of a Monet oil painting" works. "Girl, umbrella, tree-lined street, oil painting texture" does not, because the model has to guess how the pieces relate. Tag-style prompts carried over from older diffusion models are the single most common reason Seedream 4.5 results look generic.

State the purpose

If the image has a job, say so. "Design a logo for a gaming company" or "an e-commerce hero image" tells Seedream 4.5 what layout conventions to follow. ByteDance's example contrasts "Design a logo for a gaming company... The company name "PITBULL" is written on it" with "An abstract image of a dog holding a controller, and the word PITBULL on it". The first gets a usable logo; the second gets an illustration with some text.

Keep it short and specific

ByteDance notes that Seedream 4.5 understands prompts better than Seedream 3.0 and no longer produces washed-out images, so piling on words like "ultra-detailed, masterpiece, 8K, best quality" adds little. Spend your words on concrete details instead: materials, counts, positions and colors.

Text-to-Image Prompt Examples

Each example below follows the formula. Copy it, swap the subject and text, and keep the structure.

Product photography

Product shots reward precise materials and lighting. Name the surface, the light direction and the camera angle.

Commercial product photo of a matte black ceramic coffee mug on a walnut desk, thin steam rising from the rim, an open notebook and a brass pen beside it. Soft window light from the left, warm morning tones, shallow depth of field, camera at a 45-degree angle, product centered.

E-commerce hero image for a frosted glass skincare serum bottle with a gold dropper cap, standing on a wet pale-pink stone slab with small water droplets. Clean studio lighting with a soft top light and subtle reflections, minimal pastel background, square composition with empty space on the right for copy.

Posters and typography

Seedream 4.5's typography upgrade is the reason many teams pick it. Put each line of copy in quotes, say where it goes and describe the font.

Minimalist concert poster on an off-white background. A single black line drawing of a saxophone sits slightly right of center. The title "NIGHT JAZZ" is set in a large bold serif font at the top left. Below it, a smaller line reads "Live at Blue Room, Nov 14". Two lines of small sans-serif text at the bottom left list the venue address and ticket website.

E-commerce promotion banner in a modern flat style. Top title: "Today's Hot Sale". A pair of white running shoes floats in the center with a soft shadow. Below the product, the price "$59.99" in bold orange numerals. Background is a vertical gradient from coral pink to light lavender.

Portraits

For people, describe identity cues, wardrobe, expression and the lens or lighting setup you would use on a real shoot.

Editorial portrait of a woman in her thirties with short curly hair, wearing an oversized camel wool coat, standing on a rainy city street at dusk. She looks slightly past the camera with a calm expression. Neon shop signs blur in the background, cinematic color grading, 85mm lens look, soft rim light on her hair.

Studio headshot of a smiling middle-aged chef in a white double-breasted jacket, arms crossed, plain charcoal background. Classic three-point lighting, sharp focus on the eyes, natural skin texture, vertical framing from the chest up.

Infographics and knowledge visuals

ByteDance highlights that Seedream 4.5 can turn knowledge into dense visuals such as diagrams and formulas. Use precise technical terms and name the layout.

Clean educational infographic titled "How Coffee Is Made", showing five steps from left to right: harvesting, processing, drying, roasting and brewing. Each step has a simple flat icon and a one-line caption below it. Earthy brown and cream color palette, white background, horizontal layout.

Image Editing Prompts

Seedream 4.5 edits an image when you pass it as a reference and describe the change. The official guide groups edits into four operations: addition, deletion, replacement and modification.

The four edit types

The rule that matters most is to name what must not change. ByteDance's example is "Dress the tallest panda in pink Peking Opera costume and headgear, keeping its pose unchanged", not "Put that one in pink clothes".

Edit examples

Replace the plain gray wall behind the woman with a bright botanical greenhouse full of tropical plants. Keep her face, pose, hair, clothing and the lighting on her skin unchanged.

Remove the coffee cup and the phone from the table. Fill the space with the same wood grain and keep everything else in the photo exactly as it is.

Translate all English text on this poster into Spanish, matching the original font style, size and position. Do not change the illustration or the background.

Use visual markers for hard-to-describe areas

When a region is hard to describe in words, draw on the reference image first. The official guide supports arrows, bounding boxes and doodles: "Insert a TV where the red area is marked and a sofa where the blue area is marked" or "Change the word in the red highlighted area to blue italic text". This is often faster than writing a long spatial description.

Multi-Image Prompts

Multi-image work is where Seedream 4.5 earns its keep for brand and catalog teams. There are two modes: several images in, and a set of images out.

Reference-based generation

When you want to keep a character, style or product from a reference, name the reference target first and then describe the new scene. ByteDance calls these two parts the "reference target" and the "generated scene description".

Based on the character in the reference image, create a collectible vinyl figure standing on a round clear base on a wooden desk. Keep the character's hairstyle, outfit colors and facial features. Soft indoor light, realistic product photo, figure on the left third of the frame.

Multi-image fusion

With several references, refer to each by its position in image_urls: Image 1, Image 2 and so on. State what you take from each one.

Dress the model in Image 1 with the jacket from Image 2 and the sneakers from Image 3. Keep the model's face, pose and the studio background from Image 1. Match the lighting so the new clothing looks naturally photographed.

Place the perfume bottle from Image 1 on the marble pedestal from Image 2, and apply the warm film color grading of Image 3 to the whole scene. Keep the bottle's label text readable.

Image sets and storyboards

To trigger a series, ask for "a set", "a series" or a specific number of images. Seedream 4.5 keeps characters and style consistent across the set, which suits storyboards, comic panels and brand kits.

Using this logo, create a set of four brand visuals for an outdoor coffee brand named "TRAILHEAD": a paper cup, a canvas tote bag, a cap and a shop window sticker. Forest green and cream colors, simple modern style, consistent logo placement across all four images.

On APIMart, batch output requires at least one reference image, so this prompt is paired with a logo in image_urls and n set to 4.

Common Mistakes and Fixes

When text still looks wrong

Shorten the copy, enlarge it ("large bold title"), and render at 4K. If a single word is still off, run an edit pass on the result: "Change the word in the red box to "LAUNCH", keeping the font and color unchanged."

When the subject drifts across a set

Add a stable description of the character or product to every set prompt, and pass a clean reference image. Seedream 4.5 is good at consistency, but it needs an anchor.

How to Run Seedream 4.5 on APIMart

APIMart exposes Seedream 4.5 through POST https://api.apimart.ai/v1/images/generations, following the APIMart Seedream 4.5 docs. Calls are asynchronous: you get a task_id, then fetch the result.

Text-to-image request

curl --request POST \
  --url https://api.apimart.ai/v1/images/generations \
  --header 'Authorization: Bearer <token>' \
  --header 'Content-Type: application/json' \
  --data '{
    "model": "seedream-4.5",
    "prompt": "Minimalist concert poster on an off-white background. The title \"NIGHT JAZZ\" is set in a large bold serif font at the top left.",
    "size": "3:4",
    "resolution": "2K",
    "n": 1
  }'

Editing and batch request with references

curl --request POST \
  --url https://api.apimart.ai/v1/images/generations \
  --header 'Authorization: Bearer <token>' \
  --header 'Content-Type: application/json' \
  --data '{
    "model": "seedream-4.5",
    "prompt": "Using this logo, create a set of four brand visuals for an outdoor coffee brand named \"TRAILHEAD\".",
    "image_urls": ["https://example.com/trailhead-logo.png"],
    "size": "1:1",
    "resolution": "4K",
    "n": 4,
    "sequential_image_generation": "auto",
    "sequential_image_generation_options": { "max_images": 4 }
  }'

Key parameters

Getting the result

The submit call returns task_id with status submitted. Poll GET https://api.apimart.ai/v1/tasks/{task_id} until status is completed, then read the image link from data.result.images[].url (task status docs). Generated links expire after about 24 hours, so download anything you want to keep.

FAQs

When was Seedream 4.5 released?

ByteDance Seed released Seedream 4.5 in December 2025. Its official model ID, seedream-4-5-251128, reflects a late-November 2025 build, and it remains available on APIMart as seedream-4.5.

What is the best prompt structure for Seedream 4.5?

Write subject, action and scene in one natural sentence, then add style, lighting and composition, and finish with any text in double quotes. This matches ByteDance's official guidance and keeps prompts short enough that the model does not drop details.

How do I get accurate text in Seedream 4.5 images?

Put the exact words in double quotes, say where they go and describe the font, for example the title "NIGHT JAZZ" in a bold serif font at the top left. Keep the amount of copy small and use 4K for posters with small print.

How many reference images does Seedream 4.5 support?

The official BytePlus API accepts up to 14 reference images, and inputs plus outputs cannot exceed 15. On APIMart, reference images plus the number of output images (n) must also total 15 or fewer.

Can I write Seedream 4.5 prompts in languages other than English?

ByteDance lists Chinese and English as the supported prompt languages for Seedream 4.5. Write prompts in one of those two, and put text in any other language inside quotes if you need it printed; test non-Latin scripts before relying on them.

How do I edit only part of an image?

Name the exact object, describe the change and list what stays the same, for example "Replace the gray wall with a greenhouse, keeping her face, pose and clothing unchanged". For complex regions, draw a box or arrow on the reference image and refer to it in the prompt.

What is the maximum resolution of Seedream 4.5?

On the official API it outputs 2K or 4K, with custom sizes up to 4096x4096 pixels in total. On APIMart you choose 2K or 4K plus an aspect-ratio preset.

How much does Seedream 4.5 cost on APIMart?

It is billed per generated image, and batch jobs are charged for the images actually produced. See the Seedream 4.5 model page for current prices and membership discounts.

What to Watch Next

ByteDance has already moved on to the Seedream 5.0 family, and Seedream 5.0 Pro shares much of the same prompt logic, so the formula here transfers well. For a side-by-side view of the family, read our Doubao Seedream and Seedance comparison, and if you want to see how a different model handles the same principles, the FLUX 3 prompting guide is a useful companion. Save the prompts that work for you as templates, change one slot at a time, and you will get predictable Seedream 4.5 results much faster.