Seedream 5.5
Seedream 5.5 (bytedance-seed/seedream-5-5) is ByteDance Seed's proprietary image generation and editing model, released November 9, 2026 (OpenRouter listing November 12, 2026). It is a closed-weight, API-only endpoint distinguished by pixel-level preservation across up to 100 chained edit rounds, native RGBA transparent asset output, multi-reference-to-image fusion (up to 10 reference images via OpenRouter), and substantially improved small/multilingual text rendering. Architecture, parameter count, VAE/text encoder and resolution matrix are officially undisclosed; no standardized third-party benchmark scores have been published for this exact variant.
Architectural Profile & Capabilities
Seedream 5.5 is the fifth-generation flagship of ByteDance Seed's Seedream image family and the direct successor to Seedream 5.0 Pro. It is distributed exclusively as a hosted API (BytePlus ModelArk / Volcano Engine under the marketed product name "Dola Seedream 5.5", plus OpenRouter at slug bytedance-seed/seedream-5-5) — there are no open weights and no peer-reviewed technical report, arXiv paper, or reproducible evaluation table as of mid-November 2026. Capability surface. The model supports four request shapes per BytePlus ModelArk documentation: text-to-image, single-image-to-image editing, multi-reference-to-image generation, and iterative multi-turn editing. BytePlus additionally exposes an interactive local-editing mode with sketch-and-anchor control, a layer-decomposition operation (splitting a generated image into multiple compositing layers within the base image coordinate system), and a maximum output-image count of 2 per request for the layered path. OpenRouter documents up to 10 reference images per request on the hosted endpoint. ByteDance's launch demonstration chains 100 successive editing rounds — each output re-fed as the next input — and claims pixel-level preservation of non-edited regions with reduced color drift; that claim is a vendor characterization, not an independently reproduced metric. Native RGBA / transparency. Seedream 5.5 generates genuine alpha-channel output (fully or partially transparent PNG/RGBA) rather than requiring a post-hoc background-removal pass — a meaningful operational difference for design-compositing and layered-asset pipelines, and the single most defensible differentiator versus GPT Image 2.6 and Nano Banana 2.1 class models. Text rendering. ByteDance reports internal text-corruption rates of 0.4% severe and 1.7% minor on structured/dense text across Chinese, English and small type. These are vendor-reported internal evaluations; test-set size, annotation protocol and the severe/minor boundary are not published, so they are not directly comparable to a GenEval/DPG-Bench/OneIG-Bench style score. Evaluation posture. ByteDance evaluated Seedream 5.5 on MagicArena with experienced image-generation raters across six professional production categories (e-commerce, marketing, film & television, gaming, office productivity, fashion design), reporting improvements in both quality score and usability rate over Seedream 5.0 Pro in all six, with a maximum +55.62% GSB (good/same/bad) preference delta for realism in image-to-image comparisons. These are pairwise human-preference deltas against a named baseline, not standardized benchmark scores, and this dossier does not encode them as benchmarks. No published numbers exist for GenEval, DPG-Bench, OneIG-Bench, T2I-CompBench, ImgEdit, GEdit-Bench, RISE, LMArena, or Artificial Analysis image arenas for the exact variant. A single low-confidence aggregator listing (DesignArena Elo 1252) surfaced without documented comparison counts, confidence intervals, endpoint version or prompt distribution and is therefore excluded as unverifiable. Pricing (OpenRouter, single provider, no multi-provider routing). US$0.02 per standard image output, US$0.035 per high-resolution image output, US$0.008 per input image. BytePlus advertises a starting rate of US$0.02 per image. The $0/$0 per-1M-token telemetry on file reflects that this endpoint bills per-image rather than per-token; the 128K context figure describes the prompt/reference token budget, not a billed token stream. Architecture status. ByteDance has not disclosed whether the generator is a diffusion transformer, a hybrid autoregressive-plus-diffusion stack, or another design; parameter count, dense-vs-MoE composition, latent/VAE formulation, text encoder identity, and training scale are all unpublished. Any statement that Seedream 5.5 is a specific DiT with a specific active-parameter count should be treated as speculation. The only architecture facts that are operationally safe are behavioral: it is a native multi-image-input, multi-image-output model with iterative edit state carried implicitly through the request history rather than an exposed KV/context cache. Speed. OpenRouter's live leaderboard places Seedream 5.5 33rd of 62 image models by generation speed — mid-pack rather than a speed-optimized tier. No vendor or independent median latency, p95, time-to-first-byte, or throughput-at-resolution figure has been published. High-resolution tier requests should be budgeted at roughly 2–4x standard-tier wall time, and multi-turn chains (up to 100 rounds) accumulate cost and latency linearly with round count.
ByteDance has published no technical report, arXiv paper or model card for Seedream 5.5. Confirmed: no open weights; API-only via BytePlus ModelArk ("Dola Seedream 5.5") and OpenRouter; inputs accept text plus up to 10 reference images per request (OpenRouter limit); outputs are raster images including native RGBA with alpha channel; per-request maximum output image count of 2 for the layer-decomposition path; supports text-to-image, single-image editing, multi-reference fusion and iterative multi-turn editing (vendor-demonstrated to 100 chained rounds). Not confirmed / must not be asserted: dense-vs-MoE structure, total and active parameter counts, latent or VAE formulation, text/multimodal encoder identity, attention mechanism (MLA/GQA/etc.), test-time reasoning or planning tokens, exact resolution matrix (a "high-resolution" OpenRouter price tier exists but native 4K output is unverified), and whether the reference-image cap is a model constraint or an API-wrapper constraint. No exposed KV/context cache — multi-turn edit state is carried by re-submitting prior outputs as inputs, which is why cost and latency scale with round count.
Recommended Workloads & Primary Use Cases
- •No standardized, third-party benchmark scores exist for this exact variant. Only vendor-reported MagicArena GSB pairwise preference deltas (max +55.62% vs Seedream 5.0 Pro on image-to-image realism) and internal text-corruption rates (0.4% severe / 1.7% minor) are published. Do not substitute Seedream 4.0, 4.5, 5.0 or 5.0 Pro GenEval/DPG-Bench/arena numbers as proxies — they are different models with different architectures.
- •Pixel-level preservation across 100 rounds is a vendor claim, not an independently reproduced metric. In practice, drift compounds non-linearly in high-frequency regions (fine text, repetitive patterns, faces at small scale) and in chains that repeatedly re-encode through different crop/resolution tiers. Keep round counts as low as the task allows, pin a fixed output tier across the chain, and archive round-0 plus round-N for diffing.
- •Resolution and aspect governance is undocumented. A "high-resolution" billing tier exists but no published width-height matrix, no confirmed native 4K, and no stated maximum pixel count. Aspect-ratio and size parameters must be discovered empirically per provider; unsupported combinations will silently snap, letterbox, or error depending on the wrapper — always validate dimensions on the returned asset, not on the request.
- •Per-image billing plus per-input-image billing plus no multi-provider routing. OpenRouter charges $0.008 per supplied input image in addition to output cost, so a 10-reference fusion request plus iterative re-submission of prior outputs can multiply effective cost several-fold versus the headline $0.02. OpenRouter forwards all traffic to a single upstream provider — no failover, no price competition, and a single queue-priority regime.
- •Weak prompt/tool contract for programmatic use. There is no function calling, no structured tool interface and no token-level streaming of a reasoning trace; the 128K context is a prompt-and-reference budget, not a stateful conversation memory. Multi-turn behavior must be reconstructed client-side by re-uploading the prior image, and BytePlus layer-decomposition output is capped at 2 images per request. Also note NSFW/moderation policies differ between BytePlus and OpenRouter and can reject requests silently in batch jobs.
Calls to Seedream 5.5 are routed through strict Zero Data Retention inference channels. Prompts and outputs are never stored, indexed, or monitored by ByteDance.
More Models from ByteDance
Compare other engines available in this laboratory.
Seed 2.1 Turbo is ByteDance Seed's lower-latency, cost-optimized member of the Seed2.1 productivity family — a closed, multimodal (text + image, 262,144-token context) model tuned for coding, tool use, and long-horizon agent workflows at $0.50/$2.50 per 1M tokens.
ByteDance's dedicated coding model (released Feb 14, 2026 on Volcano Engine; Jul 30, 2026 on OpenRouter).
General-purpose tasks, balanced reasoning and cost