Seedream 5.5

Seedream 5.5: precise edits, transparent assets, and layer split

Seedream 5.5 is ByteDance Seed’s image model, announced on November 9, 2026. Official positioning emphasizes more stable multi-turn edits (up to 100 rounds), native RGBA transparent or semi-transparent assets, and clearer dense type. This page is about running that model here: gptimage2.asia is a third-party credit workspace, not the official 即梦 app, not BytePlus Lumina, and not ByteDance’s own console. Generation uses Seedream 5.5, not the OpenAI Images API.

Updated 2026-11-13

What it is good at

Choose Seedream 5.5 when you need pixel-stable multi-turn edits, PNG assets with an alpha channel, or a still that holds lighting, materials, and small type. It is a different stack from GPT Image 2.6 on this site: generation uses Seedream 5.5, not the OpenAI Images API. Resolution is 1K, 1.5K, or 2K (default 2K). Transparent output requires PNG.

Identifiers & options

Product name
Seedream 5.5
Provider
ByteDance Seed
Text to image
seedream/5-5-text-to-image
Image to image
seedream/5-5-image-to-image (1–10 public JPEGs, PNGs, or WEBPs, ≤10 MB each)
Layer decomposition
seedream/5-5-layer-decomposition
Aspect ratios
1:1, 4:3, 3:4, 16:9, 9:16, 2:3, 3:2, 21:9 (image-to-image also supports auto)
Size / format
1K, 1.5K, 2K (default 2K); PNG or JPEG; background auto, transparent, or opaque

Why use it here

Stable multi-turn editing

Seedream 5.5 is designed to preserve unedited regions at the pixel level across successive edits, reducing color drift and detail loss for up to 100 rounds on the same image.

Native transparent backgrounds

Request a transparent background with PNG output to get an alpha channel—no separate background-removal step for full or partial transparency.

Clearer type and more natural stills

The model targets more natural lighting, materials, and skin, and more accurate complex characters and densely arranged small text in posters and infographics.

Three ways to run it

This site is a third-party credit workspace, not the official 即梦 app, not BytePlus Lumina, and not ByteDance’s own console. Generation uses Seedream 5.5, not the OpenAI Images API. The sections below add the acceptance checks, credit rules, and caps that are easy to miss. For a task-by-task contrast, see Seedream 5.5 vs GPT Image 2.6.

Credits on this site

Live Seedream FAQ (not repeated in full on the pricing intro):

  • 1K and 1.5K cost 2 credits; 2K costs 4.
  • Reference edits also charge each input image.
  • Layer split charges the input plus each output, up to the max you set.
  • No separate Stripe list price for Seedream.

Failed generations are not charged, according to the site-wide pricing intro (re-checked 2026-11-12).

The batch generate footer currently says “Other models cost 2 credits per image,” which conflicts with Seedream 2K at 4 credits and with layer-split billing. See the callout on Pricing. Until that is edited, follow the Seedream page’s 2 / 4 ladder for Seedream jobs.

Because default size is 2K (4 credits), starting at 1K or 1.5K while you explore is the cheaper way to iterate — when the unit system itself is clear. Pack cards still say “generations” while these jobs cost 2 or 4 credits; this page does not promise a pack’s image count.

How to get a real transparent PNG — and how to check it

Official Seedream materials treat native RGBA output as a headline feature: transparent or semi-transparent assets with an alpha channel, including softer edges, so you do not have to cut the subject out later. On this site the controls are concrete:

  1. Open text-to-image or edit.
  2. Set output format to PNG and background to transparent. JPEG cannot carry an alpha channel; the live FAQ says transparency is disabled for JPEG.
  3. In the prompt, say that you need a transparent background (and whether you also need semi-transparent edges, such as hair, smoke, or glass). Ask for no backdrop, no checkerboard, and no fake ground shadow if you plan to composite the asset later.
  4. Download the original PNG from the site. Do not screenshot the preview, do not export JPEG, and do not rename a .jpg to .png.

Then check that the file is actually transparent:

  • Open it in an editor and look at the mode: you want RGBA, not RGB.
  • Place the same file on a dark, light, and strongly colored (for example red or magenta) background. Look for white fringes, a leftover boxed backdrop, a painted checkerboard, or holes that did not punch through.
  • A checkerboard in the preview is not proof. Some UIs draw a checkerboard under transparent pixels; a model can also draw a checkerboard into the picture. Only the file’s alpha channel counts.

If the result is opaque, confirm you really downloaded the PNG and that background was not left on opaque or auto. Then generate again. This site does not promise perfect hair or glass on the first try; official write-ups also treat glass and leftover fill as something to inspect.

Layer-split outputs are a special case: the live hint says the format control “applies to the base image. Separated layers are always returned as PNG.”

Layer split on this site (and the cap that differs from other docs)

Layer decomposition “splits one image into a base image plus isolated PNG layers.” On this site you can:

  1. Open Try layer split.
  2. Upload one image (required). Supported types follow the page: public JPEG, PNG, or WEBP, ≤10 MB.
  3. Optionally write a prompt that names what to separate (for example a title and a product). You can also turn on text recognition in layers.
  4. Set max output images to 2–10 (default 6). The live page says this cap exists so credit cost stays predictable.
  5. Wait (about one to three minutes) and download the base plus layers.

Why this cap matters: some third-party API documents describe a much higher max_images ceiling (for example a 2–100 range with a default of 30 on one provider’s docs). Community write-ups of official demos sometimes mention dozens or even around a hundred layers. This workspace is not those APIs. Here the ceiling is 2–10, default 6. Credits scale with the input plus each output, up to the max you set, so raising the cap raises the potential charge.

Use layer split when you want parts you can rearrange later (titles, stickers, subjects). It is not a substitute for a fully editable design document, and it does not produce SVG.

Multi-turn edits: reuse the result as the next reference

The live FAQ: you can keep editing the same image by downloading or reusing a result as the next reference. Seedream 5.5 is designed to keep unedited regions stable across many rounds, up to 100. That “up to 100” figure comes from ByteDance’s description, which this page reports; it is not a test log from this site.

A practical loop here:

  1. Generate or edit once at 1K or 1.5K if you are still exploring.
  2. Check what must stay (product shape, face, already-correct text).
  3. Open edit mode, attach the latest good output (plus any style or identity references you still need), and change one thing.
  4. Repeat, always feeding the last accepted frame, rather than starting from the first original every time unless you intend to.

Expected wait is again about one to three minutes per job. This is not a face-swap tutorial; only use photos you have the right to edit.

Seedream 5.5 examples

80 public cases compiled in the community catalog Awesome Seedream 5.5. Open a card to preview the image; when the source published a prompt, you can copy it or try it in this site’s generator.

Images and prompts come from the cited sources and the catalog at xianyu110.github.io/awesome-seedream-5.5. Figures and reviews are taken from those sources, not independently verified here. Prompts are included only when the original published them.

Try it on this site

  1. Step 1

    Open the generator

    Use the buttons on this page. The workspace opens with Seedream 5.5 already selected, including the mode you picked.

  2. Step 2

    Set Seedream options only

    You will see 1K / 1.5K / 2K, PNG or JPEG, and background controls. GPT Image 2.6-only options such as 4K, GIF, or SVG are not shown, because Seedream 5.5 does not support them.

  3. Step 3

    Generate with your credits

    Sign-in, guest trial, downloads, and gallery saving follow the same flow as other models. Credits follow this site’s 2/4 ladder: 1K or 1.5K costs 2 credits, 2K costs 4. Edits also charge each reference image; layer split charges the input plus each allowed output.

How it sits next to other models

Seedream 5.5 (this page)

ByteDance Seed model for generation, reference edits, transparent PNGs, and layer split, at up to 2K.

GPT Image 2.6

OpenAI Flare / Sunburst on this site’s Images API path. Different vendor, and the only place here that follows the GPT Image size/format controls.

Seedream 5.0 Pro

The previous Seedream line. ByteDance’s own MagicArena evaluation reports quality and usability gains for 5.5 across six use cases; see the note below.

ByteDance reports, on its MagicArena evaluation versus Seedream 5.0 Pro, higher quality and usability scores across six use cases (e-commerce, marketing, film and television, gaming, office productivity, and fashion design); usability rates in film and television and office productivity of more than twice the previous version; a pairwise preference (GSB) of +55.62% for realism in image-to-image tasks; and severe / minor text corruption rates of 0.4% and 1.7%. These figures come from ByteDance’s own evaluation, not from an independent test on this site.

FAQ

Is Seedream 5.5 actually connected, or only described here?

It is connected. Selecting Seedream 5.5 in the generator sends text-to-image, image-to-image, or layer-decomposition jobs to Seedream 5.5 and returns the finished images on this site.

Can I keep editing the same image?

Yes. Download or reuse a result as the next reference image. Seedream 5.5 is designed to keep unedited regions stable across many rounds, up to 100.

How do transparent backgrounds work?

Set output format to PNG and background to transparent. JPEG cannot carry an alpha channel, so transparency is disabled for JPEG.

What is layer decomposition?

A Seedream 5.5 mode that splits one image into a base image plus isolated PNG layers. This site lets you cap how many images come back (2–10, default 6) and optionally describe what to separate.

Does this page list a price?

On-site billing uses this site’s existing credit system: Seedream 1K and 1.5K cost 2 credits, 2K costs 4. Reference edits also charge each input image; layer split charges the input plus each output, up to the max you set. We do not quote a separate Stripe price. Pack cards still say “generations” while these jobs cost 2 or 4 credits; this page does not promise a pack’s image count.

Is this official 即梦 or BytePlus?

No. Official product pages are linked as sources. This is a third-party frontend with its own credits, caps, and wait times.

Related