# Seedream 5.0 Lite Prompt Guide: Multi-Reference, Text and 4K, Tested

Canonical page: https://fuser.studio/articles/seedream-prompt-guide

How to prompt Seedream 5.0 Lite: sentence briefs, quoted text, Figure 1/2/3 multi-reference edits, aspect ratio in the prompt and 2K–4K output, with real test outputs.

[All guides](/articles) · [Seedream in Fuser](/models/seedream)

**Quick answer:** prompt Seedream 5.0 Lite with plain sentences that name the subject, what it is doing, where it is, and the look you want. Put any words that must appear in the image inside "double quotes", say the aspect ratio in the prompt ("vertical 4:5 ad"), and when you use reference images, refer to them as Figure 1, Figure 2 and so on, and say what must stay unchanged. That is ByteDance's own Seedream prompt guide ([BytePlus ModelArk](https://docs.byteplus.com/en/docs/ModelArk/1829186)), written for Seedream 4.0 and 4.5 and linked from the Seedream 5.0 API reference, and the main points below were checked with real Seedream 5.0 Lite runs through the same fal endpoint Fuser calls.

## What Seedream 5.0 Lite is good at

Seedream 5.0 Lite is ByteDance Seed's image model released in February 2026. ByteDance describes it as a unified model that "thinks" before drawing: it reasons about the instruction, reads reference images more closely than earlier versions, and handles letters, numbers and times inside complex prompts ([ByteDance Seed](https://seed.bytedance.com/en/blog/deeper-thinking-more-accurate-generation-introducing-seedream-5-0-lite)). The same model does text-to-image and editing.

In practice that means three strengths worth prompting for:

- **Multi-reference composition.** Combine a product, a scene and a logo from separate images in one pass.
- **Short text in the image.** Titles, labels, dates and prices, spelled as written.
- **Large output.** 2K by default in Fuser, with 3K and 4K options.

ByteDance also notes the limits of a smaller model: "room for improvement in structural stability, realism, and aesthetics". Our 4K test below shows what that looks like.

## Write sentences, not tag lists

ByteDance's guide says to describe **subject + action + environment** in coherent natural language, then add style, colour, lighting or composition if they matter, and it gives a keyword list as the pattern to avoid. It also says Seedream 4.0 and later need less description than Seedream 3.0, so concise and precise beats stacking adjectives.

![Two Seedream 5.0 Lite harbour images: a keyword prompt gives a generic pink dawn with a standing figure; a sentence prompt gives a fisherman coiling rope in teal and rust fog.](https://statics.fuser.studio/cms/6b51648c-83f8-41e9-b183-3e1f0e65e316)

_Left: a keyword list. Right: the same idea as a sentence brief. Seedream 5.0 Lite, auto_2K, one run each, 28 September 2026._

We ran both forms. "fisherman, boat, harbor, dawn, fog, film photo" produced a competent but generic dawn: a pink sky, a figure standing in the bow, nothing happening. The sentence version named the action (coiling rope), the camera position (from the dock, face turned away), the medium (35mm film, fine grain) and the palette (muted teal and rust), and every one of those showed up.

A useful order:

1. **Subject** with the details that matter: "an old fisherman in a rust-red jumper".
2. **Action**: "coiling rope on the deck".
3. **Setting and time**: "a small wooden boat in a foggy harbour at dawn".
4. **Camera**: "medium shot from the dock".
5. **Look**: "shot on 35mm film, soft grey light, muted teal and rust".
6. **Purpose and format**, if it is for something specific: "a vertical 4:5 ad with space at the top for a headline".

State the purpose when you have one. ByteDance's example is "Design a logo for a gaming company..." rather than "an abstract image of a dog holding a controller": the model makes layout decisions based on what the image is for. Keep prompts under 600 English words; ByteDance warns that very long prompts scatter the model's attention and details get dropped ([ByteDance API reference](https://docs.byteplus.com/en/docs/ModelArk/1541523)). The same reference lists Chinese and English as the prompt languages for 5.0 Lite.

## Put text in double quotes

The single most reliable text tip in ByteDance's guide: wrap every word that must appear in the image in double quotation marks, as in Generate a poster with the title "Seedream 4.5".

![Two Seedream 5.0 Lite jazz posters from the same prompt: with the text in double quotes the subtitle sits on one line in small caps; without quotes it splits onto two highlighted lines.](https://statics.fuser.studio/cms/18a65770-0287-40d0-a2a2-2109c439c3eb)

_Identical prompts except for the quotation marks. Both came back at 1728 × 2304, the 3:4 the prompt asked for._

Our test used one poster prompt twice, changing only the quotes. With quotes, the title and the line "Friday 14 November · Doors 8 pm" were set exactly as written, including the middle dot, and the subtitle stayed on one line in small caps as asked. Without quotes, the words were still spelled correctly but the model treated the subtitle as loose copy, broke it onto two lines and boxed it in highlights. The quoted version is the one you could send to print.

Keep quoted text short: a title, a date line, a label or a price. For long paragraphs, add the copy afterwards in Fuser's Compositor rather than asking the model to set it.

## Reference images: name them Figure 1, 2, 3

Connect images to the Seedream node and it switches from text-to-image to editing automatically. The fal endpoint Fuser uses accepts up to 10 reference images per run ([fal model page](https://fal.ai/models/fal-ai/bytedance/seedream/v5/lite/edit)); ByteDance's own API lists 14, so 10 is the working limit in Fuser.

With more than one image, the prompt has to say what to take from each. ByteDance's pattern is to refer to each input by position and name the element: replace the character in Image 2 with the character from Image 1, in the style of Image 3. Seedream's own launch post uses "Figure 1, Figure 2" the same way. Two further rules from the guide:

- **Say what to keep.** "Keeping its pose unchanged" beats hoping. Avoid pronouns such as "that one".
- **Say what to extract.** For a reference, name the part you want (the character design, the product material, the art style), then describe the new scene.

The workflow in the image at the top of this page is a real run. Three references went in: an amber perfume bottle, a terrace with lemon trees and a SOLENE wordmark. The prompt was: "Place the perfume bottle from Figure 1 on the limestone wall of the terrace in Figure 2, in the same late-afternoon light, with a soft shadow falling to the right. Print the wordmark from Figure 3 on the front of the bottle in gold foil. Keep the bottle's shape, amber glass and brass cap unchanged. Vertical 4:5 advertising layout with clear sky at the top for a headline."

Seedream kept the bottle's shape and cap, printed SOLENE in gold on the glass, used the terrace and light, left sky at the top, and returned exactly 4:5 (1792 × 2240). The one miss: the shadow fell toward the camera, following the sun in the scene rather than the prompt. When a detail fights the physics of the reference, expect the reference to win.

## Edits: change one thing, pin the rest

For single-image edits, ByteDance lists four operations: add, remove, replace and modify. Write one concise instruction and then list what stays fixed.

We sent the finished ad back into a second Seedream node with: "Change the time of day to blue hour just after sunset: deep blue sky, the sea darker, a warm lamp glow from the left lighting the bottle. Keep the bottle, the SOLENE label, the lemon trees and the composition unchanged." The bottle, label, trees and framing held. Seedream read "lamp glow from the left" as a reason to add a lantern at top left, which is reasonable but was not asked for. If you only want light, say "warm light from the left, no visible lamp".

When a region is hard to describe, ByteDance suggests marking it on the image itself with an arrow, a box or a scribble and referring to the mark in the prompt: "Insert a TV where the red area is marked".

## Aspect ratio and resolution

The Seedream node in Fuser has an **Image Size** control with Auto, Auto 2K, Auto 3K, Auto 4K and six fixed presets: Square HD, Square, Portrait 4:3, Portrait 16:9, Landscape 4:3 and Landscape 16:9. Auto generates at 2K.

With the auto sizes, Seedream picks the shape from your prompt. ByteDance's API reference says so directly: set the resolution level and "describe its aspect ratio, shape, or purpose in the prompt". Our runs confirmed it. The poster prompt that said "vertical 3:4" came back at 1728 × 2304, the "vertical 4:5" ad at 1792 × 2240, and a landscape scene with no ratio stated came back at 2848 × 1600 (16:9). If the shape matters, write it in the prompt or pick a fixed preset.

Seedream 5.0 Lite outputs between 2560 × 1440 and 4096 × 4096 in total pixels ([fal schema](https://fal.ai/models/fal-ai/bytedance/seedream/v5/lite/text-to-image)). Smaller presets are scaled up to fit that range: Square HD gave us 3072 × 3072, not 1024 × 1024.

![Seedream 5.0 Lite auto_4K watchmaker's bench at 5504 by 3040 pixels, with 100 percent crops of a handwritten CAL. 7750 label and the watch movement.](https://statics.fuser.studio/cms/38b9a537-69fb-4f1e-9897-c56f6bd40dd5)

_Auto 4K returned 5504 × 3040. The quoted label holds up at 100%; the movement looks right at a glance but is not a real mechanism._

At Auto 4K a "wide 16:9" prompt returned 5504 × 3040. The handwritten label "CAL. 7750" is crisp at 100%. The watch movement has convincing metal and jewels but the parts do not form a working mechanism. That is the "structural stability" limit ByteDance mentions: 4K gives you more pixels, not more accuracy. Use it for print and crops, and check technical subjects at full size.

## Seed, batches and search: what not to expect

- **Seed does not lock the image.** fal's schema for Seedream 5.0 Lite lists no seed input, and in our test the same prompt with seed 1234 returned a different photo on a second run. Treat every run as new and keep outputs you like.
- **One image per run in Fuser.** ByteDance's API can return a related set from one prompt ("a series", "four storyboard images"); the Fuser node returns one image. For a set, run it again or duplicate the node, and reuse the same references to keep the look consistent. Our [consistent characters guide](/articles/consistent-characters-ai) covers that workflow.
- **No live web search.** ByteDance's launch post lists real-time search as a Seedream 5.0 Lite feature; the fal endpoint Fuser uses has no search option. Put current facts, prices and dates in the prompt yourself.

## Seedream 5.0 Lite prompts to copy

1. **Product ad from references:** "Place the product from Figure 1 on the marble counter in Figure 2, morning window light from the right. Keep the label and shape unchanged. Vertical 4:5 layout with empty space at the top."
2. **Event poster:** "A vertical 2:3 gig poster. Title "NIGHT SWIM" in bold condensed letters at the top, "Saturday 6 December · 9 pm" below it. A swimmer under blue-green water, risograph texture, two inks."
3. **Packaging set:** "Using the logo in Figure 1, design a coffee bag for a roastery called "LOW TIDE". Kraft paper, one-colour print in deep teal, front view on a white background."
4. **Style transfer:** "Apply the style of Figure 2 to Figure 1. Keep the composition and the people in Figure 1 unchanged."
5. **Outfit swap:** "Dress the person in Figure 1 in the jacket from Figure 2. Keep the pose, face and background unchanged."
6. **Sketch to render:** "Based on this floor plan, generate a photorealistic modern living room. Furniture placement must match the plan exactly. Do not include any text or pencil lines from the sketch." (adapted from ByteDance's guide)
7. **Infographic:** "A clean square infographic titled "HOW ESPRESSO WORKS" with four numbered steps, each with a simple line icon: grind, tamp, extract, serve."
8. **Relight an existing image:** "Change the lighting to overcast late afternoon. Keep every object, the framing and the colours of the products unchanged."
9. **Remove something:** "Remove the power cable on the left. Keep everything else unchanged."
10. **Editorial portrait:** "Close-up portrait of a ceramicist in a dusty studio, looking past the camera, side light from a tall window, shot on medium-format film, calm and quiet."

## Build it as a workflow in Fuser

Seedream is one node on a canvas, so the output can go straight into the next step. A common chain: product and scene references into Seedream, the result into an upscaler such as [Topaz](/models/topaz-image-upscale) for print, or into an image-to-video model such as [Kling 3.0](/models/kling-3-0-video) to animate it ([product photo to video ad](/articles/product-photo-to-video-ad)). Keep each prompt in its own text node so you can reuse it with [Gemini Image (Nano Banana)](/models/gemini-image) or [GPT Image](/models/gpt-image) and compare: [Seedream vs Nano Banana](/articles/seedream-vs-nano-banana) runs that comparison. For other editing models, see the [best AI image editing models](/articles/best-ai-image-editing-models); for text-heavy work, the [best models for text in images](/articles/best-ai-models-for-text-in-images).

## Seedream 5.0 Lite prompt cheat sheet.

What to write, and what we saw when we tested it.

### Prompting

| Goal | Write | Tested result |
| --- | --- | --- |
| Clear scene | Subject, action, setting, camera and look in sentences. | Sentence brief hit action, angle and palette; keyword list stayed generic. |
| Exact text | Put each word in "double quotes" and keep it short. | Quoted title and date line set exactly, layout as asked. |
| Several references | Figure 1, Figure 2… plus what to take from each and what stays unchanged. | Bottle, terrace and wordmark combined in one 4:5 ad. |
| Local edit | One instruction, then list what must not change. | Relit to blue hour; bottle and framing held, a lamp was added. |
| Aspect ratio | State it in the prompt with an auto size, or pick a fixed preset. | "vertical 3:4" returned 1728 × 2304. |
| High resolution | Choose Auto 3K or Auto 4K. | Auto 4K returned 5504 × 3040. |
| Same image again | Keep the output; the seed does not reproduce it. | Seed 1234 gave a different image on a second run. |

## Questions, answered.

### Which Seedream version does Fuser use?

Seedream 5.0 Lite. The Seedream node calls fal's Seedream 5.0 Lite text-to-image endpoint, and switches to the Seedream 5.0 Lite edit endpoint when you connect images.

### How many reference images can Seedream 5.0 Lite use?

Up to 10 in Fuser, which is the limit of the fal endpoint it runs on. ByteDance's own API lists up to 14. Refer to them in the prompt as Figure 1, Figure 2 and so on.

### Can Seedream 5.0 Lite generate 4K images?

Yes. Total output ranges from 2560 × 1440 to 4096 × 4096 pixels. In our test Auto 4K returned 5504 × 3040 for a wide 16:9 prompt. Fuser's default Auto setting generates at 2K.

### How do I get accurate text in Seedream?

Put the exact words in double quotation marks, keep them short, and say where they go and in what style. In our same-prompt test, quoted text followed the requested layout more closely than unquoted text.

### How do I control the aspect ratio in Seedream 5.0 Lite?

With the auto sizes, write the ratio or format in the prompt, for example "vertical 4:5 ad"; Seedream sets the width and height from it. Or pick a fixed preset such as Landscape 16:9 on the node.

### Does the seed make Seedream results reproducible?

Not on the endpoint Fuser uses. fal's schema for Seedream 5.0 Lite has no seed input, and our repeat run with the same seed produced a different image. Keep the outputs you want to reuse.

## Compose with Seedream on one canvas.

Connect references, write the brief and send the result to upscaling or video without leaving Fuser.

[Open Seedream in Fuser](https://fuser.studio/models/seedream) · [Explore all guides](https://fuser.studio/articles)

## More articles

- [Grok Imagine Image 2.0 Guide: Prompts, Editing and Settings](https://fuser.studio/articles/grok-imagine-image-guide.md)
- [Nano Banana Prompt Guide: Generate, Edit and Combine Images with Gemini](https://fuser.studio/articles/nano-banana-prompt-guide.md)
- [Best AI Models for Text in Images (2026)](https://fuser.studio/articles/best-ai-models-for-text-in-images.md)