Added the "Kling Unified Format" section (zh & en in sync; existing Kling Format docs untouched): 15 video models across the kling-* / kling-v-* series, all behind two endpoints — submit on POST /v1/videos, then poll and fetch the direct link on GET /kling-video/v1/videos/{task_id} in one call
The Models Guide page covers all 8 model pairs with capabilities, parameter tables and three-language examples; the Submit Task page splits examples into dedicated sections (t2v / i2v / first-last frame / multi-shot / motion control / element & voice references / multipart upload) that are directly reachable from the outline
New Elements & Voices page: creating / querying / deleting assets in the two namespaces (kling-element / kling-voice, kling-v-element / kling-v-voice), plus quantity and audio constraints when referenced in video generation (asset IDs are series-scoped)
Key notes: audio defaults to off (the model's cheapest tier), billing follows the actually generated tier, failed tasks are never billed; result links expire — download them promptly after generation
The endpoint quick reference gains an entry for this format, and the model naming table adds the kling-* / kling-v-* prefixes
The image-to-image "Doubao Seedream Format" section is updated against the 5.0 series (zh & en in sync): doubao-seedream-5-0 joins the verified list, image now accepts multiple reference images (string / string[]), and sequential_image_generation is marked as unsupported by doubao-seedream-5-0
New "Seedream Layer Decomposition" section: doubao-seedream-5-0-pro and doubao-seedream-5-0-flash-260915 pass layer_decomposition: true to output a base image plus independent layers, up to 17 images, with three-language examples
The "Image generation" row of the endpoint quick reference and the capability list now mention Seedream layer decomposition
Field-tested correction on the "Query & Download" page (zh & en in sync): content_url is a signed temporary URL — token-free but time-limited, and expired URLs return 502. The earlier "long-lived, safe to store in a database" wording was wrong and now reads "download and re-host immediately; do not distribute as a long-term link"
Adds the content-endpoint auth rule: GET /v1/tasks/{task_id}/artifacts/{key}/content without ?access=requires a token; only the artifacts[].content_url form that already carries ?access= is token-free
The submit page error table gains 502 Artifact upstream authentication failed (expired image URL signature), and the polling examples now carry re-host reminders
The single Image Generation page is split into "Text-to-Image & Image-to-Image / Image Editing / Vega Format" (zh & en in sync), with sub-pages following the unified "Endpoint → Authentication → Request Body → …" skeleton, restoring the right-side outline
New Vega async task pages (new version): POST /v1/tasks/vega for submission (text-to-image and image-to-image share the endpoint — omit input for t2i), GET /v1/tasks/{task_id} for polling (QUEUED / IN_PROGRESS / SUCCESS / FAILURE), and /artifacts for download; content_url is a token-free direct link on this site (timeliness corrected in the entry above)
Billing rules: 0 pre-charged at submission, settled on actual usage at completion, failed tasks not billed; pro image-to-image includes the first 3 reference images free with +10,000 tokens each from the 4th on; generation measured at about 19 s for Lite and 98 s for Pro
The pre-migration synchronous integration (t2i Vega series + i2i Vega format) is preserved as the "Legacy Guide": the sync endpoint no longer supports Vega (tiered expr run failed), noted prominently at the top with a pointer to the async API; the endpoint quick reference gains a Vega async row
#2026-10-12 · Doubao Video Docs: Field Corrections
camera_fixed moves from the "not supported" list into the formal parameter table (zh & en in sync): re-tested 2026-10 on t2v — submission returns 200 and the parameter takes effect upstream (true / false show clearly different camera movement), correcting the earlier "returns 400" finding; frames / service_tier remain unsupported
New i2v material compliance warning: image / images do not accept photos of real people — Volcano's compliance check rejects them with 400 InputImageSensitiveContentDetected.PrivacyInformation (no task created, no billing); use illustration / anime / product / scenery images for the first frame
The music section is reorganized by vendor (zh & en in sync, mirroring the video section): the Volcengine Format directory holds "New (Async Tasks)" and "Legacy (Sync API)" — the new flow submits via POST /volcengine/music/submit and polls via GET /volcengine/music/fetch/{task_id}, with models doubao-music-bgm (instrumental) / doubao-music-song (song) and statuses queued / processing / completed / failed; the legacy synchronous /v1/audio/speech (volcengine-bgm / volcengine-song) keeps working for existing users
The submit page documents the full metadata reference: BGM duration 30–120 s / Song 30–240 s, strict genre / mood / timbre enums (English values, case-sensitive; non-enum values return 502 with the full legal list), input word count 5–2000, and input becoming optional once metadata.lyrics is provided
Billing rules: billed by audio duration, pre-charged on the requested duration at submission and settled against the actual upstream duration on completion (difference refunded or charged); failed tasks are not billed and refunded in full; data.url is a signed temporary URL — download and transfer promptly
MiniMax Music chat streaming now has its own "MiniMax Format" page (same vendor-page pattern as the video section), with the Suno deprecation note kept there; the endpoint quick reference music row now points to the async submit endpoint
#2026-09-29 · Video Docs: MiniMax-H3 & Material Modes
The video section gains a "MiniMax Format" page (zh & en in sync): MiniMax-H3 reuses the generic task endpoint /v1/video/generations with all three modes — text-to-video, image-to-video, video-to-video. The gateway rejects /v1/chat/completions for this model (guards against mis-billing as text tokens)
Live-tested calibration: only 768p / 2K resolutions (omitting it fails validation against 720p), metadata.ratio required for text-to-video, duration validated as an integer from 4 to 15 (default 5); asset URLs must be reachable from the upstream network (overseas image hosts fail with unreachable 2013), and failed tasks are not charged
Query response adds the channel mirror fields: data.data.output.video_url (upstream OSS link, expires in ~24 hours) while result_url is this platform's proxied download link; the video overview TIP, endpoint quick reference, and status-page NOTE note this channel
The generic task page (Create Video Generation Task) now documents the full set of input material modes: first/last frame (metadata.first_frame_url / last_frame_url), multi-reference-image (multiple images), video-to-video and multi-reference-video (metadata.video_urls), reference audio (metadata.audio_urls), plus a mode table, material format constraints, and the "metadata merged as a whole / top-level fields silently dropped" rule; three new scenario examples added (zh & en in sync)
The Kling section is split into "New (all models)" and "Legacy (v2.5-turbo only)" directories. New pages are rewritten for the TokenHub direct plugin — field table (resolution / audio / multi_shot replace the rich mode tokens), model × parameter matrix, Kling-envelope responses (with final_unit_deduction / created_at mirrors), element & voice references moved to metadata.contents, and the callback_url tunnel replaced by direct contents / options
The new Submit Task page notes that legacy syntax (mode tiers, native settings object, metadata tunnels) keeps working unchanged; mode: "mock" returns 400 (blocked to avoid real charges)
New "Motion Control" page (motion-control models): metadata.contents assembly (exactly one character image + one motion video, optional prompt/element), required character_orientation and its effect on the motion video duration cap, resolution / audio settings, asset constraints and verified behavior (720p tier at 1264×720, output length follows the reference video)
The legacy section is condensed into a single usage guide: only the field table and minimal examples v2.5-turbo actually uses, dropping ultra / asset tokens / callback_url content to reduce noise
Video documentation reorganized into "task overviews + integration formats": top-level create / query overview pages (with POST / GET method badges), and four format directories — Doubao / Kling / Jimeng / Sora — each split into Submit / Query sub-pages, replacing the single video.mdx
Sub-pages follow a unified "Endpoint → Request Body → Request Examples → Response Body → Response Example" skeleton, restoring the right-side outline; examples unified to cURL / Python / Node.js, with both text-to-video and image-to-video samples for Doubao
Doubao (Seedance) calibrated against live tests: duration via seconds, frame via metadata, the silent-drop list for top-level fields; a per "model × mode" parameter matrix recorded (mini i2v fixed at 720p, resolution unsupported, etc.)
Sidebar: format directories collapsed by default; method badges green for GET / blue for POST (Tag override in theme/index.tsx)
Kling / Jimeng / Sora query pages add the unified task JSON response body and examples; the endpoint quick reference video row now shows /v1/video/generations
Audio documentation split by capability into three standalone pages — TTS / Music / STT — replacing the single audio.mdx page
Added the MiniMax Music chat streaming API: images passthrough documented; the Suno format is marked deprecated but kept for reference
Video docs: added tables and examples for element / voice management (zh & en in sync)
#2026-09-16 · Video & Audio Field Testing Calibrations
Kling: resolution now determined by mode (std 720P / pro 1080P); added native / nomulti switches and the v2.6 native forced-1080p note
Kling: asset URL restrictions rewritten from live testing (Tencent Cloud direct links, failing hosts and recommended order); image-to-video constraints added (jpg/png, ≤50MB, ≥300px on each side)
Removed the "Quick Start" module: both hero buttons on the home page and the quickstart directory were removed; related links now point to in-product entry points
Flattened product routes: /products/api-model/... → /tokenhub/...; the home page now enters product docs only through the four product cards
Each product page now shows only its own sidebar, no longer listing other products
The "Products" navbar item became a dropdown with the four products
FAQ and Changelog pages hide the sidebar (single pages need no navigation)
Renamed the API Model Service directory/route from api-model to tokenhub (display name and external platform domains unchanged)
Full API Model Service documentation: Introduction, Quick Start, API Key Management, Billing & Top-up, Base URLs & Auth, API Calls (text / audio / image), Error Codes, Models (categories / billing / naming / choosing), App Integration (Cherry Studio / Claude Code / Codex / others / CC Switch)
Top-level FAQ in four categories: account / API / models / clients
Placeholder pages for YUAN, STMind, and the Compute Scheduling Platform