From script and references to a reviewed final cut

Shortest workflow

1. Sign in and open the AIGC Workshop.
2. Choose an image or video task and an available model.
3. Set prompt, dimensions, duration, resolution, or reference media.
4. Submit and wait for a final state in Task records.
5. Preview and download your result.

Beginner drama: a complete script to server delivery

1. Complete import: paste text or upload UTF-8 TXT, MD, or DOCX up to 10 MB (10 MiB). All extracted text is stored separately and can be downloaded as UTF-8 text. DOCX reads main-document paragraphs and tables, not images, comments, or executable objects. Long manuscripts are partitioned automatically at chapter or paragraph boundaries. Oversized input is rejected, never silently shortened.
2. Sequential analysis: select an available text group and model, then analyze the current unit or continue the entire manuscript. Each step saves characters, locations, props, episodes, and shots; source ranges must cover the complete unit. Pause after the current step. Completed steps are skipped, and uncertain requests use the same original identity and complete response, including its real finish reason.
3. Style and continuity: AI proposes 3–5 script-derived visual directions with one recommendation. Refine with a sentence. Later units continue the selected style and confirmed character, location, and prop identities and reference assets from the manuscript. Changed descriptions or styles require media review.
4. Shots and sound: select image, video, and speech groups and models. Generate references in batches, adopt and publish them, then produce shots. Shot cards also support continuing individual unfinished or outdated pictures and videos. Successful results are retained; uncertain batches are checked first. Measured speech drives dialogue and caption alignment.
5. Editing and delivery: returning from advanced editing preserves shot order, trims, music, and manually edited captions. Longer episodes render in segments of no more than 60 seconds. Once all current segments are exported, join them on the existing server into one downloadable episode without an external editor.

Files and extracted text support up to 10 MiB; DOCX main XML is bounded at 40 MiB and 1024 ZIP entries. Each production unit contains up to 6000 UTF-16 units and a 1 MiB editable document. Unit projects are created lazily; one manuscript consumes one ordinary project slot. Each user supports 20 manuscript and ordinary-project slots combined and 200 MiB of manuscript storage. A unit supports 12 episodes and 64 entities; an episode supports 32 scenes, 64 shots, 64 spoken lines, and 300 planned seconds.

Normal segment renders remain limited to 60 seconds, 720p, and 30 fps. Episode assembly supports up to 300 seconds and 128 MiB, submitted by the project owner using their own complete, ordered, current segments from the same project and episode. Incompatible codecs fail explicitly. Final renders and assembly share a daily limit of 10 jobs, at most 2 active jobs, and existing queue and storage admission. Model charges use base prices, the selected group multiplier, and actual usage; budgets control precharge admission. Import, saves, captions, mixing, rendering, and assembly reuse the existing server without another paid rendering subscription. Completed analysis means saved planning; artistic quality, character performance, lip sync, and word alignment still depend on actual model capabilities and review.

Drama: source material to reviewed delivery

Work through five production stages. Generated results remain candidates until reviewed and adopted. Recheck affected media after script, reference, dialogue, or edit changes.

1. Writing and storyboard: confirm a project bible and episode outline for adaptations and original stories. Existing scripts can go straight to normalization. Select an available text group and model for each stage, preview, then adopt. Each episode retains its shots, selected media, canvas, and timeline.
2. Characters and visuals: define characters, locations, props, and style. Generate reference images from descriptions or upload media. Adopt and attach images, then publish an immutable reference version. Multiple references and first/last frames are separate modes, enabled only when supported by the selected model.
3. Dialogue and sound: assign each speaker a speech group, model, supported voice ID, and speed. Generate individual lines or attach uploaded audio, listen, then adopt. Text or voice changes invalidate the previous take. Native video audio is a separate route to avoid duplicate dialogue.
4. Previsualization and editing: assemble an animatic from storyboard images and a rough cut from selected videos. Planned, requested, and measured durations stay separate. Align adopted speech to create dialogue cues and sentence captions from measured audio lengths. Realign after trimming or reordering. Correct native-audio captions manually.
5. Review and checks: validate sources, reference versions, assets, trims, and speech, then review story, continuity, sound/captions, and editing/cover. Owners can require approval: save the revision, render a preview, approve that revision, then export and deliver. Later edits require approval of the new revision.

Batch production and costs

Select shots and confirm the budget before submitting image or video jobs, up to 32 nodes per batch. Results return as candidates for their original shots. Independent failed shots do not block others; retry the failed subset. Check uncertain submissions before starting a new batch. Budgets control precharge admission; actual model charges use base prices, the selected group multiplier, and actual usage.

Current render limits

Previews, captions, mixing, and final composition run on the existing server with no desktop editor or additional paid rendering service. Each segment render supports up to 60 seconds, 720p, 30 fps, 64 clips, 64 dialogue cues, 2 background audio tracks, and 128 captions; complete episode segments can be joined into one file of up to 300 seconds. Measured sentence timing does not imply word alignment, automatic lip sync, or artistic approval.

Models and groups come from live capabilities

Select an available group, then a model supported by that group for text, images, video, or speech. A model needs an enabled channel and the exact endpoint capability to enter its selector.

Use the live image catalog and live video catalog for the public machine-readable view.

Image editing and speech APIs

Use POST /v1/images/edits with multipart/form-data: model, prompt, image, and optional mask, in a group that supports editing. Workshop edits save a separate result. Speech uses POST /v1/audio/speech with model, input, voice, response_format, and optional speed. Select a model that supports speech synthesis; Studio can archive MP3 or WAV.

API clients use http://localhost:3000 and their account's Bearer key. Workshop session routes and project archival use a different interface from ordinary API calls.

Image generation

Use /v1/images/generations. The Direct base URL is recommended for long responses. Use a regular key in an image group supported by the model. Edit any retired smart media key in API Keys first.

Video and multiple reference assets

Video generation is a long-running task and should use the Direct API. Seedance uses a regular video group supported by the model.

{
  "prompt": "Use @Image 1 and speak with @Audio 1",
  "duration": 15,
  "ratio": "16:9",
  "resolution": "720p",
  "references": [
    {"media_type":"image","role":"reference_image","url":"YOUR_PUBLIC_IMAGE_URL","alias":"Image 1"},
    {"media_type":"audio","role":"reference_audio","url":"YOUR_PUBLIC_AUDIO_URL","alias":"Audio 1"}
  ]
}

Aliases in the prompt must match aliases in references. The generation service must be able to fetch each URL.

Reference video upload

Upload a local reference video with POST /api/aigc/uploads, then submit the complete signed data.url from the response. The upload layer checks the extension, declared MIME, and file signature; the video limit is 64 MiB. Keep the exp and sig query parameters.

curl -X POST 'http://localhost:3000/api/aigc/uploads' \
  -H "Authorization: Bearer $DONGAPI_API_KEY" \
  -F 'file=@reference.mp4'

# Use the returned data.url
{"references":[{"media_type":"video","role":"reference_video","url":"SIGNED_REFERENCE_URL","alias":"Video 1"}]}

• MiniMax-H3/qy: a separate reference-video contract, available only when declared in the live catalog. The current MiniMax-H3 contract excludes video references.
• Seedance: reference-video counts and roles come from the selected group's video_profile_by_group. With a reference video, Seedance 2.5 per-second models also bill the reference video's length by the official rules (see billing). Ark asset:// IDs are not supported; upload the file and use the returned URL.
• Models without declared video-reference support hide that upload control; unsupported requests return field errors before billing.

The selected group's request profile determines the controls. A successful upload does not mean every video model supports reference video; refresh expired signed URLs.

Grok video billing and MiniMax-H3 parameters

Grok images are billed per image: each image costs the configured base price times the group multiplier. Grok videos follow the live catalog price_unit: second charges the base rate for the selected mode and resolution × submitted duration × the current group multiplier. Single-image and 2–4-image generation use their image-generation rate; reference images carry no separate count fee. Version 1.5 without images uses the text-only rate. 1080p requires version 1.5 with reference images. Base rates and group multipliers remain separate; discounts never rewrite the base rate. Grok second rates follow the platform pricing rule: the USD numeric rate forms the CNY base at 1 USD = 1 CNY, then the group multiplier applies once. The official CNY reference uses approximately 1 USD = 7 CNY for comparison only. For example, a base rate of CNY P per second for 8 seconds at multiplier 0.6 costs P × 8 × 0.6 = 4.8P CNY; the current rate of each mode and resolution is on the Grok model page. Mode, resolution, duration, and rate freeze at submission and appear in Usage logs. Models still marked request and historical tasks retain their original per-request contract. The live model catalog is authoritative for currency and current prices.

The current MiniMax-H3 request contract is 768P, 5–15 seconds, default 5. Text generation ratios are 21:9, 16:9, 4:3, 1:1, 3:4, and 9:16. Image/audio references and first/last-frame controls depend on the selected group's live profile. This contract does not include video references. The current exact MiniMax-H3 uses the configured per-request base price and group multiplier; the separate qy contract uses duration-based billing. Check the selected model’s live price unit.

• Frames use first_frame_url / last_frame_url, or references with first_frame / last_frame roles. Do not mix frames with ordinary multiple references.
• Ordinary references use matching media_type and role values. Counts come from the selected group's video_profile_by_group.
• MiniMax-H3/qy has a separate contract. Its video-reference, callback, and watermark controls do not apply to other models.
• The current Grok video route uses a compatibility contract. Official multi-image, voice, keyframe, editing, or extension capabilities require support in the selected group before controls are enabled.

Parameter catalog and advanced controls

The authenticated pricing catalog provides video_profile_by_group with duration_options, aspect_ratios, resolution_options, reference_roles, modes, and parameters. Single videos, drama shots, batch jobs, and canvas use the same group-specific contract. Advanced settings show declared controls only. Omitted values use model defaults; explicit false and 0 remain in requests.

Recover the original request when acceptance is uncertain. Changing models, groups, settings, or references requires a separate task. Studio video references use signed URLs suitable for long jobs; permission changes still revoke access. Refresh expired asset URLs.

Task records and downloads

Image and video task records remain available for at least 24 hours. A user can inspect and download only results associated with their own account.

API routes and success signals

A successful task has a final success state, a preview/download result, and a matching request ID and charge in Usage logs.

Poll and download with the same account

Create with POST /v1/videos, retain the returned public task ID, then poll GET /v1/videos/{task_id}. Status progresses through queued / in_progress to completed or failed. Submission alone does not mean the video is ready. After completion, download with GET /v1/videos/{task_id}/content. API clients must send their 懂api Bearer token for both polling and download; the task URL alone does not grant access.

curl 'http://localhost:3000/v1/videos/TASK_ID' \
  -H "Authorization: Bearer $DONGAPI_API_KEY"

curl 'http://localhost:3000/v1/videos/TASK_ID/content' \
  -H "Authorization: Bearer $DONGAPI_API_KEY" \
  --fail --output video.mp4