Seedream 5.5: ByteDance's Image Model with Transparent Output, 99-Layer Decomposition, OCR and Pricing

What Seedream 5.5 is

Seedream 5.5 is an image generation model from ByteDance, offered on BytePlus ModelArk as Dola Seedream 5.5 with the model ID dola-seedream-5-5-261028. BytePlus lists it among its October 2026 releases and calls it ByteDance's latest image creation model; this site dates it 28 October 2026, from the model ID. It does text-to-image, single-image-to-image and multi-reference-to-image generation, plus interactive editing, using the same requests as Seedream 5.0 pro and flash with only the model name changed. BytePlus names three improvements: more consistent multi-turn editing, native transparent images and finer layer decomposition. On LMArena's text-to-image leaderboard Seedream 5.5 has no LMArena rating.

Multi-turn editing and transparent images

Seedream 5.5 is built for editing one image over several turns, keeping the subject, composition and untouched areas consistent; each turn takes the previous output as its input and states what to change and what to keep. For the most consistent results BytePlus suggests an input without an alpha channel and an opaque background while editing. The model can also produce RGBA images with a real alpha channel, from a text prompt alone or from one reference image or up to 10, for example to cut stickers out of photos onto a transparent background. Transparent output needs the background set to transparent and PNG as the format; JPEG output cannot be transparent.

Layer decomposition with text recognition

With layer decomposition switched on, Seedream 5.5 splits one image into a base image and up to 99 separate layers, against 16 for Seedream 5.0 pro. Layers always come back as transparent PNGs. It can also return rich-text OCR for ordinary text layers: the text itself, the position of each line and approximate styles such as a similar font family, so a design tool can rewrite, translate or restyle the text. BytePlus notes the limits: artistic and handwritten text and non-text layers return no OCR, font matches are approximate, and the editing itself is done by the application, not the model. Each request reserves image quota for its maximum layer count and returns the unused part.

Sizes, formats and limits

Output can be set as a tier, 1K, 1.5K or the default 2K, or as exact pixels from 921,600 up to about 4.62 million in total. Images return as a URL or Base64, in PNG by default or JPEG, with an optional watermark that is on by default. Inputs can be image URLs or Base64 data. Generated image URLs are kept for 24 hours and then deleted, so files need to be saved promptly. Rate limits are counted in images per minute per model and version.

Pricing

BytePlus bills Seedream 5.5 per image. An output image costs $0.020 when its pixel count is up to 1536 squared times 1.05 squared, roughly 1.5K, and $0.035 up to 2048 squared times 1.05 squared, roughly 2K. Each input image costs $0.008. In layer decomposition each layer is billed at the tier of its own size, and images blocked by content moderation are not charged. That makes Seedream 5.5 cheaper per output image than Seedream 5.0 pro, which BytePlus prices at $0.045 and $0.09, though 5.0 pro charges $0.003 for input images after the first, which is free.

Official Resources