bytedance-seedream

What it does

Tells an agent to build and operate production image generation and natural-language image editing with ByteDance Seedream through first-party Ark image-generation APIs (China) or ModelArk (global). It includes model and region selection, multi-reference and grouped outputs, streaming, secure artifact handling, retries, cost controls, prompting, safety, and rights review.

How it works

The skill directs the agent to select among documented deployment options and corresponding API bases and model IDs, using the proper host depending on China vs global needs. It requires constructing requests to POST {API_BASE}/images/generations with a JSON body including fields like model, prompt, and optional image, while enforcing input formats, sizes, and response formats. It specifies handling of grouped outputs, streaming, and, for certain models, optional web_search tool usage. It also instructs parsing of streaming events such as image_generation.partial_succeeded, image_generation.partial_failed, and image_generation.completed, and details about artifact acquisition (downloading URL responses within a 24-hour window or decoding b64_json). It emphasizes strict handling of API keys, region residency caveats, and avoiding disallowed parameters like seed or guidance_scale unless explicitly documented in the endpoint contract. It includes a production workflow with steps for policy envelope, input validation, prompt construction, job ledger, narrowly scoped retries, artifact safety checks, and post-release evaluation.

When to use it

Use when an application calls Seedream directly through the first-party Ark image-generation API for text-to-image, reference-guided generation, natural-language edits, multi-image fusion, grouped outputs, and streamed grouped outputs. It is not for Seedance video generation, third-party gateways, consumer apps, model training, or UI automation.

What it can touch

The skill references what to send and how to interpret responses to endpoints like POST {API_BASE}/images/generations and handling of fields such as model, prompt, image, size, response_format, watermark, and output_format. It requires careful handling of tokens and uses of API bases corresponding to deployment surface: China (Volcengine Ark) and global (BytePlus ModelArk). It also mentions tools and interactions only at the level of documented streaming and potential web_search usage for 5.0 Lite, with outputs including URLs or base64 JSON.

Caveats

Notes include that model IDs and hosts are not interchangeable across deployments, endpoint selection is not a data-residency guarantee, and some features (like seed or guidance_scale) are not documented for current endpoints. It warns to recheck model catalog, availability, pricing, and retirement pages before production. It also highlights that some claims are provider claims and should be validated against representative prompts, and that there are strict limits on grouped outputs, streaming, and artifact handling depending on model version.