Seedream 5.5 Prompts: 12 to Copy, From Stickers to Edits
Twelve Seedream 5.5 prompts you can paste today: product shots, transparent stickers, posters with exact text, multi-reference characters and edit chains that keep the rest intact. Plus what three real test runs on each::labs showed.
By Eftal Yurtseven · Nov 16, 2026 · 9 min read

The first image from a new model is easy to love. The fifth edit is where you find out whether you can work with it. Good Seedream 5.5 prompts are built for that fifth edit: they say what changes, what stays, and what the background should be made of, so the model doesn't have to guess. ByteDance's Seed team released Seedream 5.5 on November 9, and as of today it runs on each::labs as Seedream 5.5 Text to Image and Seedream 5.5 Edit.
The overview lives in Seedream 5.5: 100 Edit Rounds and Native Transparency. This is the practical follow-up: 12 prompts to paste, the phrasing behind them, and what we saw when we ran three of them on each::labs this morning.
The Building Blocks of Seedream 5.5 Prompts
Every prompt below uses the same five parts. Subject first, as a noun with its material ("a frosted glass serum bottle with a brushed gold cap"). Setting second: the surface, the backdrop, or for a cutout, the explicit absence of one. Light third, and a named source ("a single softbox from the left") beats any adjective. Medium fourth: product photography, risograph print, flat vector, vinyl sticker. Text last, only when you want letters, inside quotation marks.
Then come the controls that sit outside the prompt. On each::labs both endpoints take resolution (1K, 1.5K or 2K, default 2K), aspect_ratio (auto, 1:1, 16:9, 9:16, 4:3, 3:4, 3:2, 2:3 or 21:9) and output_format (jpeg by default, or png). The Edit endpoint adds image, a list of 1 to 10 reference URLs. The schema recommends prompts under 600 English words. Ours sit between 40 and 90.

Product Shots: Name the Surface, the Light and the Lens
Describe only the product and you get a generic gradient and flat light. Give it a set instead.
Prompt 1 (hero shot). A frosted glass serum bottle with a brushed gold dropper cap, standing on a slab of honed travertine. Single softbox from the upper left, crisp shadow falling to the right, a thin rim of light along the glass edge. Pale sand backdrop, shallow depth of field, shot on a 100mm macro lens. Clean commercial product photography, no text on the bottle.
Prompt 2 (lifestyle shot). The same style of frosted glass serum bottle on a bathroom shelf at 8 AM, morning sun through a ribbed glass window throwing soft stripes across the tiles. A folded linen towel and a sprig of eucalyptus beside it, slightly out of focus. Warm, quiet, editorial mood, photographed at eye level.
Run the first at 1:1 for a catalog grid and the second at 3:4 for a feed. Prompt 1 bans text on purpose: if you want a bare bottle for your own label later, say so rather than hope.

Ask for Transparency Like You Mean It
This is the feature people will search for, so get the phrasing right. ByteDance's Seedream 5.5 documentation says the model generates native RGBA images with an alpha channel from text and from up to 10 reference images, for stickers, icons and product materials. Its own text-to-RGBA example simply ends with "Set a transparent background." The same docs note that PNG preserves the alpha channel and JPEG does not.
On each::labs that translates into two moves. Write the transparent background into the prompt, and set output_format to png, because the default is jpeg and a JPEG has nowhere to store transparency. We ran Prompt 3 this way at 1K and 1:1 and got a 1024 by 1024 RGBA PNG: about 70 percent of pixels fully transparent, the lemon and its border opaque, no checkerboard baked in.
Prompt 3 (sticker, tested). A die-cut sticker of a cheerful cartoon lemon wearing round sunglasses, thick white sticker border, glossy vinyl highlight, flat bold colors, centered with even padding. Transparent background, no shadow, no backdrop.
Prompt 4 (icon set). A set of four flat vector icons for a weather app: sun, cloud, raindrop, snowflake. Rounded strokes of equal weight, two-tone navy and coral, arranged in a 2 by 2 grid with generous equal spacing, no overlap. Transparent background, no frame, no drop shadows.
Prompt 5 (cutout from a photo, Edit endpoint). Extract the ceramic coffee mug from Image 1 and place it alone on a transparent background. Preserve its exact shape, glaze color, handle, chip on the rim and the reflections on its surface. Remove the table, the hand and everything else. Keep the edges sharp and intact.
"No shadow" matters more than it looks: a soft drop shadow is half transparent, rarely what a sticker sheet wants. Prompt 5 borrows ByteDance's own extraction structure: what to extract, what to preserve, what to remove, then clean edges.
Posters With Text: Quote Every Word, Ban the Rest
Put every word you want on the image inside quotation marks, and say where it goes and how big it should be relative to the rest. Then say that nothing else should be written. That last line is the lesson from our second test.
We ran a night market poster at 2K and 2:3 with a quoted headline, "LANTERN NIGHT MARKET", and a quoted subtitle, "Fridays 7 PM to Midnight". Both came back spelled correctly, in a clean hierarchy over a convincing risograph texture. But the model decided the poster needed more copy: three lines of small caps under the subtitle and a three-line tagline at the bottom, spelled fine and never requested. Helpful is still off-brief. Prompt 6 is the version we'd run now, with the closing line added after that test.
Prompt 6 (poster, revised after testing). Vertical poster for a night market. Headline at the top in large bold condensed sans serif: "LANTERN NIGHT MARKET". Below it, smaller text: "Fridays 7 PM to Midnight". Illustration of paper lanterns strung over food stalls, warm orange and deep teal palette, risograph print texture, generous margins, clean typographic hierarchy. No other text, letters or signage anywhere in the image.
Prompt 7 (text removal edit, tested). Remove the three small lines of text below the subtitle and the three lines of text at the bottom of the poster, and fill those areas with the surrounding sky and street. Keep the headline "LANTERN NIGHT MARKET", the subtitle "Fridays 7 PM to Midnight", the lanterns, the stalls, the people, the colors, the print texture and the composition exactly as they are.
We sent Prompt 7 to the Edit endpoint with the poster as the only reference and aspect ratio on auto. Both unwanted blocks disappeared, the sky and the street were filled in cleanly, and the result came back at the same 1664 by 2496 size. Side by side, the lanterns, the cook, the crowd and the texture looked unchanged to us. It left a short decorative rule near the bottom, which we hadn't mentioned. It did what we said, not what we meant.

Edit Instructions: Change One Thing, List What Stays
ByteDance's guidance for multi-turn editing is short and worth taking literally: state clearly what to change and what to preserve. Its own example asks the subject to blow a pink bubblegum bubble and then lists, item by item, what must survive: identity, facial features, gaze, hairstyle, earrings, clothing, pose, background, composition and visual style.
That list is the trick. "Keep everything else the same" is a wish; a named inventory is an instruction. Write the change as one sentence with one verb (add, remove, replace, recolor, move), then a preserve sentence naming what a viewer would notice if it drifted: faces, text, product shape, camera angle. Prompt 7 follows that pattern, which is why the headline survived an edit right next to it.
Multi-Reference Prompts: Number Your Images
Seedream 5.5 Edit accepts up to 10 references in one request, and ByteDance's own examples address them as "Image 1" and "Image 2". Do the same. Refer to each image by its number in the order you send them, and give each one a single role: who, what they wear, where they are, what it should look like.
Prompt 8 (character consistency). Use the character from Image 1, keeping her face, freckles, short copper hair and green eyes exactly as they are. Dress her in the yellow rain jacket from Image 2. Place her on the rainy tram platform from Image 3, waiting with a paper coffee cup. Match the soft overcast light and muted palette of Image 3. Medium shot, eye level.
Prompt 9 (character sheet). Using the character in Image 1, create a character sheet on a plain light gray background: front view, three-quarter view, side view and back view, standing in a row at the same scale. Keep her face, hair, outfit and proportions identical in every view. Even studio lighting, no text.
Run Prompt 9 before a campaign. A clean turnaround becomes Image 1 for every scene after it.

Edit Chains: One Turn, One Change
The headline promise of Seedream 5.5 is consistency across many edit turns. On each::labs a chain is simple to build: send the output URL of one Edit call as the image of the next. ByteDance's docs add a practical tip: for the most consistent multi-turn edits, use inputs without an alpha channel. Leaving output_format on jpeg mid-chain keeps every input opaque; switch to png only for a final transparent export.
Here's a three-turn chain on a photo of a sneaker on a concrete step.
Prompt 10 (turn 1). Change the sneaker's laces from white to bright orange. Preserve the shoe's shape, the white leather, the sole, the stitching, the concrete step, the lighting, the camera angle and the composition.
Prompt 11 (turn 2). Add a few fallen autumn leaves on the concrete step around the sneaker. Preserve the sneaker exactly as it is now, including the orange laces, and keep the lighting, the camera angle and the composition unchanged.
Prompt 12 (turn 3). Shift the light to late golden hour, warmer and lower from the right, with longer shadows. Preserve the sneaker, the orange laces, the leaves and their positions, the step and the framing.
Turn 2 says "including the orange laces" on purpose: each preserve list should carry the previous turn's result. Keep the aspect ratio and resolution the same for every turn, so the only thing that changes between frames is the thing you asked for.
What Our Three Seedream 5.5 Test Runs Showed
We ran three generations on each::labs for this post. The sticker proved transparency works with a prompt line plus png output. The poster showed exact spelling plus unrequested extra copy. The edit removed two text blocks and left a busy illustration looking untouched. The product, multi-reference and chain prompts follow ByteDance's documented phrasing but weren't among our runs, so tune them on your own references.
Mistakes That Waste a Generation
The most common: asking for transparency and leaving the format on jpeg. You get a nice image and no alpha channel. Second is writing "transparent" about the object when you mean the background: "a transparent glass bottle" and "a glass bottle on a transparent background" are different requests, and the docs' own acrylic standee example uses both on purpose.
Third, stacking three changes in one turn, so you can't tell which one caused the drift. Fourth, describing a reference in words instead of pointing at it by number. Fifth, leaving text uncontrolled: if a poster or a package could carry words, the model may write some.
Save every prompt that works with its settings. If you're comparing models first, our Seedream 5.5 vs GPT Image 2.6 breakdown covers where each one bets, and the text-to-image catalog sits behind the same API key. For layered files rather than flat cutouts, the Seedream v5 Layer Decomposition model is the right tool, and the whole Seedream v5 family lives on one page. Build a chain once and reuse it with each::labs flows.
Frequently Asked Questions
What makes good Seedream 5.5 prompts?
Specific nouns. Lead with the subject and its material, give it a surface and a light source, name the medium, and quote any text exactly. For edits, one change per turn plus a list of what must be preserved.
How do I get a transparent PNG from Seedream 5.5 on each::labs?
Two moves together. End the prompt with "transparent background, no shadow, no backdrop" and set output_format to png. In our test that returned an RGBA PNG with a real alpha channel. Left on the default jpeg, there's no channel to hold transparency.
How many reference images can Seedream 5.5 Edit use?
Up to 10 per request on each::labs, passed as a list in image. Refer to them in the prompt as Image 1, Image 2 and so on, in the order you send them, and give each a single role.
How do I stop Seedream 5.5 from changing things I didn't ask it to?
Name them. A preserve sentence listing the face, the text, the product shape and the framing holds far better than "keep everything else". In a chain, include the previous turn's change in that list.



